Skip to content

[Frontend] Add Comprehensive Core Aten IR Operator Support #637

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sdh1014 wants to merge 6 commits into
base: main
Choose a base branch
from
Open

[Frontend] Add Comprehensive Core Aten IR Operator Support #637

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
fd4b774
[fronted]: Add comprehensive Core Aten IR operator support
sdh1014 Dec 2, 2025
1ab66fd
[tests] Add comprehensive Core Aten IR operator test cases
sdh1014 Dec 2, 2025
23d89b5
[frontend] Add _embedding_bag_forward_only operator mapping
sdh1014 Dec 3, 2025
368ee84
[Frontend] Restore special broadcast pattern handling in index_op
sdh1014 Dec 3, 2025
5f94ee6
[test] standardize test_lt_scalar.py and test_le_scalar.py format
sdh1014 Dec 3, 2025
eab29ec
[fronted] Improve IndexOp broadcast handling in _index_op_all_tensors
sdh1014 Dec 3, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
209 changes: 203 additions & 6 deletions frontend/Python/frontend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -118,7 +118,6 @@ def __init__(
"view.default": ViewOp,
"ones.default": OnesOp,
"full.default": FullOp,
"lt.Tensor": LessThanOp,
"embedding.default": EmbeddingOp,
"masked_fill.Scalar": MaskedFillOp,
"slice.Tensor": SliceOp,
Expand All @@ -136,10 +135,11 @@ def __init__(
"index.Tensor": IndexOp,
"neg.default": NegOp,
"cat.default": CatOp,
"squeeze.dim": SqueezeOp,
"bmm.default": BatchMatmulOp,
"div.Tensor": DivOp,
"div.Tensor_mode": DivTensorModeOp,
"_softmax.default": SoftmaxOp,
"_log_softmax.default": LogSoftmaxOp,
"clone.default": CloneOp,
"silu.default": SiluOp,
"add.Tensor": AddOp,
Expand All @@ -157,12 +157,55 @@ def __init__(
"getitem": GetItemOp,
"convolution.default": Conv2dOp,
"max_pool2d_with_indices.default": MaxPool2dWithIndicesOp,
"max_pool1d.default": MaxPool1dOp,
"max_pool1d_with_indices.default": MaxPool1dOp,
"max_pool3d.default": MaxPool3dOp,
"max_pool3d_with_indices.default": MaxPool3dOp,
"avg_pool1d.default": AvgPool1dOp,
"avg_pool3d.default": AvgPool3dOp,
"adaptive_max_pool1d.default": AdaptiveMaxPool1dOp,
"adaptive_max_pool2d.default": AdaptiveMaxPool2dOp,
"adaptive_avg_pool1d.default": AdaptiveAvgPool1dOp,
"_adaptive_avg_pool2d.default": AdaptiveAvgPool2dOp,
"_adaptive_avg_pool3d.default": AdaptiveAvgPool3dOp,
"relu.default": ReluOp,
"iota.default": IotaOp,
"sigmoid.default": SigmoidOp,
"scalar_tensor.default": ScalarTensorOp,
"where.self": WhereOp,
"sqrt.default": SqrtOp,
"abs.default": AbsOp,
"log.default": LogOp,
"ceil.default": CeilOp,
"floor.default": FloorOp,
"maximum.default": MaximumOp,
"minimum.default": MinimumOp,
"bitwise_not.default": BitwiseNotOp,
"logical_not.default": LogicalNotOp,
"clamp.default": ClampOp,
"logical_and.default": LogicalAndOp,
"logical_or.default": LogicalOrOp,
"bitwise_or.Tensor": BitwiseOrOp,
"bitwise_xor.Tensor": BitwiseXorOp,
"amin.default": AminOp,
"avg_pool2d.default": AvgPool2dOp,
"logical_xor.default": LogicalXorOp,
"prod.default": ProdOp,
"prod.dim_int": ProdOp,
"neg.default": NegOp,
"where.self": WhereOp,
"eq.Tensor": EqTensorOp,
"ne.Tensor": NeTensorOp,
"gt.Tensor": GtTensorOp,
"ge.Tensor": GeTensorOp,
"lt.Tensor": LtTensorOp,
"le.Tensor": LeTensorOp,
"log10.default": Log10Op,
"log2.default": Log2Op,
"log1p.default": Log1pOp,
"expm1.default": Expm1Op,
"constant_pad_nd.default": ConstantPadNdOp,
"masked_fill.Scalar": MaskedFillOp,
"reciprocal.default": ReciprocalOp,
"clamp_min.default": ClampMinOp,
"clamp_max.default": ClampMaxOp,
Expand All @@ -175,21 +218,176 @@ def __init__(
"gt.Scalar": GtOp,
"_scaled_dot_product_flash_attention_for_cpu.default": ScaledDotProductFlashAttentionForCpuOp,
"ge.Scalar": GeOp,
"gt.Tensor": GreaterThanOp,
"_unsafe_index.Tensor": UnsafeIndexOp,
"eq.Scalar": EqualOp,
"copy.default": CopyOp,
"slice_scatter.default": SliceScatterOp,
"le.Tensor": LeOp,
"bitwise_and.Tensor": BitwiseAndTensorOp,
"index_put.default": IndexPutOp,
"ne.Scalar": NeScalarOp,
"cumsum.default": CumsumOp,
"eq.Tensor": EqualOp,
"cumprod.default": CumProdOp,
"sort.default": SortOp,
"_tensor_constant": TensorConstantOp,
"lift_fresh_copy.default": LiftFreshCopyOp,
"repeat.default": RepeatOp,
"as_strided.default": AsStridedOp,
"scatter.src": ScatterSrcOp,
"scatter.value": ScatterValueOp,
"scatter_reduce.two": ScatterReduceOp,
"gather.default": GatherOp,
"native_layer_norm.default": NativeLayerNormOp,
"native_group_norm.default": NativeGroupNormOp,
"_native_batch_norm_legit.default": NativeBatchNormLegitOp,
"_native_batch_norm_legit.no_stats": NativeBatchNormLegitNoStatsOp,
"_native_batch_norm_legit_no_training.default": NativeBatchNormLegitNoTrainingOp,
"native_dropout.default": NativeDropoutOp,
"upsample_bilinear2d.vec": UpsampleBilinear2dVecOp,
"upsample_nearest2d.vec": UpsampleNearest2dVecOp,
"grid_sampler_2d.default": GridSampler2dOp,
"col2im.default": Col2imOp,
"sym_size.int": SymSizeOp,
"sym_stride.int": SymStrideOp,
"sym_numel.default": SymNumelOp,
"sym_storage_offset.default": SymStorageOffsetOp,
"gelu.default": GeluOp,
"flip.default": FlipOp,
"leaky_relu.default": LeakyReluOp,
"hardtanh.default": HardtanhOp,
"elu.default": EluOp,
"sign.default": SignOp,
"round.default": RoundOp,
"trunc.default": TruncOp,
"sinh.default": SinhOp,
"cosh.default": CoshOp,
"tan.default": TanOp,
"exp2.default": Exp2Op,
"zeros.default": ZerosOp,
"zeros_like.default": ZerosLikeOp,
"ones_like.default": OnesLikeOp,
"full_like.default": FullLikeOp,
"all.default": AllOp,
"all.dim": AllOp,
"any.default": AnyOp,
"any.dim": AnyOp,
"isinf.default": IsInfOp,
"isnan.default": IsNanOp,
"floor_divide.default": FloorDivideOp,
"fmod.Tensor": FmodOp,
"fmod.Scalar": FmodOp,
"remainder.Tensor": RemainderOp,
"remainder.Scalar": RemainderOp,
"pow.Tensor_Tensor": PowTensorTensorOp,
"softplus.default": SoftplusOp,
"hardswish.default": HardswishOp,
"tile.default": TileOp,
"stack.default": StackOp,
"lerp.Tensor": LerpOp,
"lerp.Scalar": LerpOp,
"clamp.Tensor": ClampTensorOp,
"le.Scalar": LeScalarOp,
"lt.Scalar": LtScalarOp,
"index_select.default": IndexSelectOp,
"scatter_add.default": ScatterAddOp,
"arange.start_step": ArangeStartStepOp,
"min.dim": MinDimOp,
"argmin.default": ArgMinOp,
"squeeze.default": SqueezeOp,
"squeeze.dim": SqueezeDimOp,
"squeeze.dims": SqueezeDimsOp,
"unfold.default": UnfoldOp,
"topk.default": TopkOp,
"unbind.int": UnbindOp,
# Batched matrix operations
"baddbmm.default": BaddbmmOp,
# Special math functions
"lgamma.default": LgammaOp,
"digamma.default": DigammaOp,
"i0.default": I0Op,
"erfc.default": ErfcOp,
# Cumulative operations
"cummax.default": CummaxOp,
"cummin.default": CumminOp,
# Tensor clamp operations
"clamp_min.Tensor": ClampMinTensorOp,
"clamp_max.Tensor": ClampMaxTensorOp,
# Additional elementwise operations
"hypot.default": HypotOp,
"copysign.Tensor": CopysignOp,
"nextafter.default": NextafterOp,
"masked_scatter.default": MaskedScatterOp,
"rev.default": RevOp,
# Scalar arithmetic operations
"add.Scalar": AddScalarOp,
"sub.Scalar": SubScalarOp,
"div.Scalar": DivScalarOp,
"div.Scalar_mode": DivScalarModeOp,
"pow.Scalar": PowScalarOp,
# Reduction operations
"mean.default": MeanDefaultOp,
"var.correction": VarCorrectionOp,
"var.dim": VarDimOp,
"any.dims": AnyDimsOp,
# Trigonometric functions
"acos.default": AcosOp,
"asin.default": AsinOp,
"atan.default": AtanOp,
"atan2.default": Atan2Op,
"acosh.default": AcoshOp,
"asinh.default": AsinhOp,
"atanh.default": AtanhOp,
# Other operations
"fill.Scalar": FillScalarOp,
"diagonal.default": DiagonalOp,
"alias.default": AliasOp,
"empty.memory_format": EmptyOp,
"rand.default": RandOp,
"randn.default": RandnOp,
"select_scatter.default": SelectScatterOp,
"split_with_sizes.default": SplitWithSizesOp,
"max.dim": MaxDimOp,
"nonzero.default": NonzeroOp,
# Standard deviation operations
"std.default": StdDefaultOp,
"std.dim": StdDimOp,
"std.correction": StdCorrectionOp,
# Additional reduction operations
"sum.default": SumDefaultOp,
"all.dims": AllDimsOp,
"var.default": VarDefaultOp,
# Norm operations
"norm.Scalar": NormScalarOp,
"norm.ScalarOpt_dim": NormScalarOptDimOp,
# Backward operations (Gradient Computation)
"_adaptive_avg_pool2d_backward.default": AdaptiveAvgPool2dBackwardOp,
"avg_pool2d_backward.default": AvgPool2dBackwardOp,
"convolution_backward.default": ConvolutionBackwardOp,
"embedding_dense_backward.default": EmbeddingDenseBackwardOp,
"max_pool2d_with_indices_backward.default": MaxPool2dWithIndicesBackwardOp,
"native_group_norm_backward.default": NativeGroupNormBackwardOp,
"native_layer_norm_backward.default": NativeLayerNormBackwardOp,
# Bitwise scalar operations
"bitwise_and.Scalar": BitwiseAndScalarOp,
"bitwise_or.Scalar": BitwiseOrScalarOp,
"bitwise_xor.Scalar": BitwiseXorScalarOp,
# Padding operations
"reflection_pad1d.default": ReflectionPad1dOp,
"reflection_pad2d.default": ReflectionPad2dOp,
"reflection_pad3d.default": ReflectionPad3dOp,
"replication_pad2d.default": ReplicationPad2dOp,
"replication_pad3d.default": ReplicationPad3dOp,
# Other missing core aten operations
"empty_strided.default": EmptyStridedOp,
"randperm.default": RandpermOp,
# Core Aten remaining operations
"_embedding_bag.default": EmbeddingBagOp,
"_embedding_bag_forward_only.default": EmbeddingBagOp,
"_cdist_forward.default": CdistForwardOp,
"_pdist_forward.default": PdistForwardOp,
"_fft_r2c.default": FftR2cOp,
"_local_scalar_dense.default": LocalScalarDenseOp,
"resize_.default": ResizeOp,
"resize.default": ResizeOp,
}

@property
Expand Down Expand Up @@ -313,7 +511,6 @@ def _compile_fx(

def _compiler(_gm: torch.fx.GraphModule, _inputs: List[torch.Tensor]):
"""Compile a FX graph in Aten/Prims IR to MLIR."""
nonlocal params_flat
num_cached_kv = 0
if self._model_config.decode_with_cache:
num_cached_kv = self._model_config.num_hidden_layers * 2
Expand Down
Loading