TileOPs-Metax/tileops/kernels/__init__.py

218 lines
6.2 KiB
Python

from .attention import (
FlashAttnBwdPostprocessKernel,
FlashAttnBwdPreprocessKernel,
GQABwdKernel,
GQABwdWgmmaPipelinedKernel,
GQADecodeKernel,
GQADecodePagedKernel,
GQAFwdFP8Fa3ContractPtxAccBN224WsTmaVKernel,
GQAFwdKernel,
GQAFwdWgmmaPipelinedKernel,
GQAPrefillFwdKernel,
GQAPrefillPagedWithFP8KVCacheFwdKernel,
GQAPrefillPagedWithKVCacheFwdKernel,
GQAPrefillPagedWithKVCacheRopeAppendKernel,
GQAPrefillPagedWithKVCacheRopeFwdKernel,
GQAPrefillWithKVCacheFwdKernel,
GQAPrefillWithKVCacheRopeAppendKernel,
GQAPrefillWithKVCacheRopeFwdKernel,
GQASlidingWindowFwdKernel,
GQASlidingWindowFwdWgmmaPipelinedKernel,
GQASlidingWindowVarlenFwdKernel,
GQASlidingWindowVarlenFwdWgmmaPipelinedKernel,
MeanPoolingFwdKernel,
MHABwdKernel,
MHABwdWgmmaPipelinedKernel,
MHADecodeKernel,
MHADecodePagedKernel,
MHAFwdKernel,
MHAFwdWgmmaPipelinedKernel,
MLADecodeKernel,
MLADecodeWsKernel,
NSACmpFwdVarlenKernel,
NSAFwdVarlenKernel,
NSATopkVarlenKernel,
SparseMlaKernel,
)
from .bmm import BmmFp8Kernel, BmmKernel
from .convolution import (
Conv1dKernel,
Conv1dPointwiseKernel,
Conv2d1x1Kernel,
Conv2dKernel,
Conv3dKernel,
GroupConv1dKernel,
GroupConv2dKernel,
GroupConv3dKernel,
)
from .deltanet import DeltaNetBwdKernel, DeltaNetFwdKernel
from .deltanet_recurrence import (
DeltaNetDecodeFP32Kernel,
DeltaNetDecodeKernel,
DeltaNetDecodeRawCudaFlaStyleKernel,
)
from .dropout import DropoutKernel
from .elementwise import BinaryKernel, FusedGatedKernel, UnaryKernel
from .engram import EngramDecodeKernel, EngramGateConvBwdKernel, EngramGateConvFwdKernel
from .fft import FFTC2CKernel
from .fp8_lightning_indexer import FP8LightningIndexerKernel
from .fp8_quant import FP8QuantKernel
from .gated_deltanet import (
GatedDeltaNetBwdKernel,
GatedDeltaNetFwdKernel,
GatedDeltaNetPrefillFwdKernel,
)
from .gated_deltanet_recurrence import (
GatedDeltaNetDecodeFP32Kernel,
GatedDeltaNetDecodeKernel,
GatedDeltaNetDecodeRawCudaFlaStyleKernel,
)
from .gemm import GemmFp8BlockScaledKernel, GemmFp8EpilogueKernel, GemmKernel, GemvKernel
from .gemm_maca import GemmMACAKernel
from .gla import GLABwdKernel, GLAFwdKernel
from .gla_recurrence import GLADecodeFP32Kernel, GLADecodeKernel
from .grouped_gemm import GroupedGemmKernel
from .kernel_base import Kernel
from .mhc import MHCPostKernel, MHCPreKernel
from .moe import MoePermuteAlignKernel
from .norm import (
BatchNormBwdKernel,
BatchNormFwdInferKernel,
BatchNormFwdTrainKernel,
GroupNormKernel,
LayerNormKernel,
RMSNormKernel,
)
from .pool import (
AvgPool1dKernel,
AvgPool1dSpatialKernel,
AvgPool2dKernel,
AvgPool2dSpatialKernel,
AvgPool3dKernel,
AvgPool3dSpatialKernel,
MaxPool1dKernel,
MaxPool1dWithIndicesKernel,
MaxPool2dKernel,
MaxPool2dWithIndicesKernel,
MaxPool3dKernel,
MaxPool3dWithIndicesKernel,
)
from .quant_swiglu_channel_cast_transpose import QuantSwiGLUChannelCastTransposeKernel
from .rope import (
RopeLlama31Kernel,
RopeLongRopeKernel,
RopeNeoxKernel,
RopeNeoxPositionIdsKernel,
RopeNonNeoxKernel,
RopeYarnKernel,
)
from .topk_selector import TopkSelectorKernel
__all__ = [
"AvgPool1dKernel",
"AvgPool1dSpatialKernel",
"AvgPool2dKernel",
"AvgPool2dSpatialKernel",
"AvgPool3dKernel",
"AvgPool3dSpatialKernel",
"BatchNormBwdKernel",
"BatchNormFwdInferKernel",
"BatchNormFwdTrainKernel",
"BinaryKernel",
"BmmFp8Kernel",
"BmmKernel",
"Conv1dKernel",
"Conv1dPointwiseKernel",
"Conv2d1x1Kernel",
"Conv2dKernel",
"Conv3dKernel",
"DeltaNetBwdKernel",
"DeltaNetDecodeFP32Kernel",
"DeltaNetDecodeKernel",
"DeltaNetDecodeRawCudaFlaStyleKernel",
"DeltaNetFwdKernel",
"DropoutKernel",
"EngramDecodeKernel",
"EngramGateConvBwdKernel",
"EngramGateConvFwdKernel",
"FFTC2CKernel",
"FP8LightningIndexerKernel",
"FP8QuantKernel",
"FlashAttnBwdPostprocessKernel",
"FlashAttnBwdPreprocessKernel",
"FusedGatedKernel",
"GLABwdKernel",
"GLADecodeFP32Kernel",
"GLADecodeKernel",
"GLAFwdKernel",
"GQABwdKernel",
"GQABwdWgmmaPipelinedKernel",
"GQADecodeKernel",
"GQADecodePagedKernel",
"GQAFwdFP8Fa3ContractPtxAccBN224WsTmaVKernel",
"GQAFwdKernel",
"GQAFwdWgmmaPipelinedKernel",
"GQAPrefillFwdKernel",
"GQAPrefillPagedWithFP8KVCacheFwdKernel",
"GQAPrefillPagedWithKVCacheFwdKernel",
"GQAPrefillPagedWithKVCacheRopeAppendKernel",
"GQAPrefillPagedWithKVCacheRopeFwdKernel",
"GQAPrefillWithKVCacheFwdKernel",
"GQAPrefillWithKVCacheRopeAppendKernel",
"GQAPrefillWithKVCacheRopeFwdKernel",
"GQASlidingWindowFwdKernel",
"GQASlidingWindowFwdWgmmaPipelinedKernel",
"GQASlidingWindowVarlenFwdKernel",
"GQASlidingWindowVarlenFwdWgmmaPipelinedKernel",
"GatedDeltaNetBwdKernel",
"GatedDeltaNetDecodeFP32Kernel",
"GatedDeltaNetDecodeKernel",
"GatedDeltaNetDecodeRawCudaFlaStyleKernel",
"GatedDeltaNetFwdKernel",
"GatedDeltaNetPrefillFwdKernel",
"GemmFp8BlockScaledKernel",
"GemmFp8EpilogueKernel",
"GemmKernel",
"GemmMACAKernel",
"GemvKernel",
"GroupConv1dKernel",
"GroupConv2dKernel",
"GroupConv3dKernel",
"GroupNormKernel",
"GroupedGemmKernel",
"Kernel",
"LayerNormKernel",
"MHABwdKernel",
"MHABwdWgmmaPipelinedKernel",
"MHADecodeKernel",
"MHADecodePagedKernel",
"MHAFwdKernel",
"MHAFwdWgmmaPipelinedKernel",
"MHCPostKernel",
"MHCPreKernel",
"MLADecodeKernel",
"MLADecodeWsKernel",
"MaxPool1dKernel",
"MaxPool1dWithIndicesKernel",
"MaxPool2dKernel",
"MaxPool2dWithIndicesKernel",
"MaxPool3dKernel",
"MaxPool3dWithIndicesKernel",
"MeanPoolingFwdKernel",
"MoePermuteAlignKernel",
"NSACmpFwdVarlenKernel",
"NSAFwdVarlenKernel",
"NSATopkVarlenKernel",
"QuantSwiGLUChannelCastTransposeKernel",
"RMSNormKernel",
"RopeLlama31Kernel",
"RopeLongRopeKernel",
"RopeNeoxKernel",
"RopeNeoxPositionIdsKernel",
"RopeNonNeoxKernel",
"RopeYarnKernel",
"SparseMlaKernel",
"TopkSelectorKernel",
"UnaryKernel",
]