Skip to content

[Torch] add aten._scaled_mm op support and FX import plumbing #4558

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
catcor01 wants to merge 6 commits into
base: main
Choose a base branch
from
Open

[Torch] add aten._scaled_mm op support and FX import plumbing #4558

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
c4b7aef
[Torch] add aten._scaled_mm op support and FX import plumbing
catcor01 Apr 24, 2026
8052ff9
[Torch] Test _scaled_mm abstract interp
catcor01 May 6, 2026
c99c3c3
[Torch] Tighten _scaled_mm abstract interp
catcor01 May 6, 2026
12efa5c
[FX] Fix _scaled_mm ragged test shape
catcor01 May 6, 2026
98eef49
[Torch] Check _scaled_mm matrix divisibility
catcor01 May 6, 2026
ada4200
[Torch] Regenerate _scaled_mm abstract interp library
catcor01 May 7, 2026
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
30 changes: 30 additions & 0 deletions include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6623,6 +6623,36 @@ def Torch_AtenMmOp : Torch_Op<"aten.mm", [
}];
}

def Torch_Aten_ScaledMmOp : Torch_Op<"aten._scaled_mm", [
AllowsTypeRefinement,
HasValueSemantics,
ReadOnly
]> {
let summary = "Generated op for `aten::_scaled_mm : (Tensor, Tensor, Tensor, Tensor, Tensor?, Tensor?, int?, bool) -> (Tensor)`";
let arguments = (ins
AnyTorchTensorType:$self,
AnyTorchTensorType:$mat2,
AnyTorchTensorType:$scale_a,
AnyTorchTensorType:$scale_b,
AnyTorchOptionalTensorType:$bias,
AnyTorchOptionalTensorType:$scale_result,
AnyTorchOptionalIntType:$out_dtype,
Torch_BoolType:$use_fast_accum
);
let results = (outs
AnyTorchOptionalTensorType:$result
);
let hasCustomAssemblyFormat = 1;
let extraClassDefinition = [{
ParseResult Aten_ScaledMmOp::parse(OpAsmParser &parser, OperationState &result) {
return parseDefaultTorchOp(parser, result, 8, 1);
}
void Aten_ScaledMmOp::print(OpAsmPrinter &printer) {
printDefaultTorchOp(printer, *this, 8, 1);
}
}];
}

def Torch_Aten_IntMmOp : Torch_Op<"aten._int_mm", [
AllowsTypeRefinement,
HasValueSemantics,
Expand Down
16 changes: 8 additions & 8 deletions lib/Conversion/TorchToTosa/TosaLegalizeUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -297,27 +297,27 @@ std::optional<Value> getConstTensor<float>(PatternRewriter &rewriter,
(src.isF32() && dest.isInteger(8)) ||
(src.isF32() && dest.isBF16()) ||
(src.isF32() && dest.isF16()) ||
(src.isF32() && isa<Float8E4M3Type>(dest)) ||
(src.isF32() && isa<Float8E4M3FNType>(dest)) ||
(src.isF32() && isa<Float8E5M2Type>(dest)) ||
// f16 -> *
(src.isF16() && dest.isInteger(32)) ||
(src.isF16() && dest.isInteger(16)) ||
(src.isF16() && dest.isInteger(8)) ||
(src.isF16() && dest.isBF16()) ||
(src.isF16() && dest.isF32()) ||
(src.isF16() && isa<Float8E4M3Type>(dest)) ||
(src.isF16() && isa<Float8E4M3FNType>(dest)) ||
(src.isF16() && isa<Float8E5M2Type>(dest)) ||
// bf16 -> *
(src.isBF16() && dest.isInteger(32)) ||
(src.isBF16() && dest.isInteger(16)) ||
(src.isBF16() && dest.isInteger(8)) ||
(src.isBF16() && dest.isF32()) ||
(src.isBF16() && isa<Float8E4M3Type>(dest)) ||
(src.isBF16() && isa<Float8E4M3FNType>(dest)) ||
(src.isBF16() && isa<Float8E5M2Type>(dest)) ||
// fp8e4m3 -> *
(isa<Float8E4M3Type>(src) && dest.isBF16()) ||
(isa<Float8E4M3Type>(src) && dest.isF32()) ||
(isa<Float8E4M3Type>(src) && dest.isF16()) ||
(isa<Float8E4M3FNType>(src) && dest.isBF16()) ||
(isa<Float8E4M3FNType>(src) && dest.isF32()) ||
(isa<Float8E4M3FNType>(src) && dest.isF16()) ||
// fp8e5m2 -> *
(isa<Float8E5M2Type>(src) && dest.isBF16()) ||
(isa<Float8E5M2Type>(src) && dest.isF32()) ||
Expand Down Expand Up @@ -514,8 +514,8 @@ LogicalResult getConvOpsAccType(PatternRewriter &rewriter,
} else if (inputElemTy.isInteger(16) && weightElemTy.isInteger(8) &&
outputElemTy.isInteger(48)) {
accType = mlir::TypeAttr::get(rewriter.getIntegerType(48));
} else if ((isa<Float8E4M3Type>(inputElemTy) &&
isa<Float8E4M3Type>(weightElemTy) && outputElemTy.isF16()) ||
} else if ((isa<Float8E4M3FNType>(inputElemTy) &&
isa<Float8E4M3FNType>(weightElemTy) && outputElemTy.isF16()) ||
(isa<Float8E5M2Type>(inputElemTy) &&
isa<Float8E5M2Type>(weightElemTy) && outputElemTy.isF16())) {
accType = mlir::TypeAttr::get(rewriter.getF16Type());
Expand Down
4 changes: 2 additions & 2 deletions lib/Dialect/Torch/IR/TorchTypes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -191,8 +191,8 @@ static bool isValidTorchDtype(Type dtype) {
// Builtin floating point types.
if (isa<Float16Type, BFloat16Type, Float32Type, Float64Type>(dtype))
return true;
if (isa<Float8E5M2Type, Float8E4M3FNType, Float8E5M2FNUZType,
Float8E4M3FNUZType, Float8E4M3B11FNUZType, Float8E8M0FNUType>(dtype))
if (isa<Float8E5M2Type, Float8E4M3FNType, Float8E8M0FNUType,
Float8E5M2FNUZType, Float8E4M3FNUZType, Float8E4M3B11FNUZType>(dtype))
return true;

if (isa<Torch::StringType>(dtype))
Expand Down
16 changes: 16 additions & 0 deletions lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8407,6 +8407,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
" %0 = call @__torch__.torch.jit._shape_functions.mm(%arg0, %arg1) : (!torch.list<int>, !torch.list<int>) -> !torch.list<int>\n"
" return %0 : !torch.list<int>\n"
" }\n"
" func.func @\"__torch_mlir_shape_fn.aten._scaled_mm\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.list<int>, %arg3: !torch.list<int>, %arg4: !torch.optional<list<int>>, %arg5: !torch.optional<list<int>>, %arg6: !torch.optional<int>, %arg7: !torch.bool) -> !torch.list<int> {\n"
" %0 = call @__torch__.torch.jit._shape_functions.mm(%arg0, %arg1) : (!torch.list<int>, !torch.list<int>) -> !torch.list<int>\n"
" return %0 : !torch.list<int>\n"
" }\n"
" func.func @\"__torch_mlir_shape_fn.aten._int_mm\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>) -> !torch.list<int> {\n"
" %0 = call @__torch__.torch.jit._shape_functions.mm(%arg0, %arg1) : (!torch.list<int>, !torch.list<int>) -> !torch.list<int>\n"
" return %0 : !torch.list<int>\n"
Expand Down Expand Up @@ -15544,6 +15548,18 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
" }\n"
" return %6 : !torch.int\n"
" }\n"
" func.func @\"__torch_mlir_dtype_fn.aten._scaled_mm\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>, %arg2: !torch.tuple<int, int>, %arg3: !torch.tuple<int, int>, %arg4: !torch.optional<tuple<int, int>>, %arg5: !torch.optional<tuple<int, int>>, %arg6: !torch.optional<int>, %arg7: !torch.bool) -> !torch.int {\n"
" %none = torch.constant.none\n"
" %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
" %1 = torch.aten.__isnot__ %arg6, %none : !torch.optional<int>, !torch.none -> !torch.bool\n"
" %2 = torch.prim.If %1 -> (!torch.int) {\n"
" %3 = torch.prim.unchecked_cast %arg6 : !torch.optional<int> -> !torch.int\n"
" torch.prim.If.yield %3 : !torch.int\n"
" } else {\n"
" torch.prim.If.yield %0#1 : !torch.int\n"
" }\n"
" return %2 : !torch.int\n"
" }\n"
" func.func @\"__torch_mlir_dtype_fn.aten._int_mm\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>) -> !torch.int {\n"
" %int3 = torch.constant.int 3\n"
" %none = torch.constant.none\n"
Expand Down
11 changes: 10 additions & 1 deletion projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -908,6 +908,9 @@ def aten〇mv〡shape(self: List[int], vec: List[int]) -> List[int]:
def aten〇mm〡shape(self: List[int], mat2: List[int]) -> List[int]:
return upstream_shape_functions.mm(self, mat2)

def aten〇_scaled_mm〡shape(self: List[int], mat2: List[int], scale_a: List[int], scale_b: List[int], bias: Optional[List[int]] = None, scale_result: Optional[List[int]] = None, out_dtype: Optional[int] = None, use_fast_accum: bool = False) -> List[int]:
return upstream_shape_functions.mm(self, mat2)

Comment thread
catcor01 marked this conversation as resolved.
def aten〇_int_mm〡shape(self: List[int], mat2: List[int]) -> List[int]:
return upstream_shape_functions.mm(self, mat2)

Expand Down Expand Up @@ -4682,6 +4685,13 @@ def aten〇mm〡dtype(self_rank_dtype: Tuple[int, int], mat2_rank_dtype: Tuple[i
dtypes = [self_dtype, mat2_dtype]
return promote_dtypes(ranks, dtypes)

def aten〇_scaled_mm〡dtype(self_rank_dtype: Tuple[int, int], mat2_rank_dtype: Tuple[int, int], scale_a_rank_dtype: Tuple[int, int], scale_b_rank_dtype: Tuple[int, int], bias_rank_dtype: Optional[Tuple[int, int]] = None, scale_result_rank_dtype: Optional[Tuple[int, int]] = None, out_dtype: Optional[int] = None, use_fast_accum: bool = False) -> int:
Comment thread
catcor01 marked this conversation as resolved.
self_rank, self_dtype = self_rank_dtype
mat2_rank, mat2_dtype = mat2_rank_dtype
if out_dtype is not None:
return out_dtype
return self_dtype

def aten〇_int_mm〡dtype(self_rank_dtype: Tuple[int, int], mat2_rank_dtype: Tuple[int, int]) -> int:
self_rank, self_dtype = self_rank_dtype
mat2_rank, mat2_dtype = mat2_rank_dtype
Expand Down Expand Up @@ -6369,4 +6379,3 @@ def _create_argparse() -> argparse.ArgumentParser:

if __name__ == "__main__":
main(_create_argparse().parse_args())

3 changes: 3 additions & 0 deletions projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -574,6 +574,9 @@ def emit_with_mutating_variants(key, **kwargs):
# Non-elementwise tensor compute ops
emit("aten::linear : (Tensor, Tensor, Tensor?) -> (Tensor)")
emit("aten::mm : (Tensor, Tensor) -> (Tensor)")
emit(
"aten::_scaled_mm : (Tensor, Tensor, Tensor, Tensor, Tensor?, Tensor?, int?, bool) -> (Tensor)"
)
emit("aten::_int_mm : (Tensor, Tensor) -> (Tensor)")
emit("aten::addmm : (Tensor, Tensor, Tensor, Scalar, Scalar) -> (Tensor)")
emit("aten::matmul : (Tensor, Tensor) -> (Tensor)")
Expand Down
Loading
Loading