build: manually update PyTorch version (#3627)

Set PyTorch and TorchVision version to nightly release 2024-08-18.
This commit also updates the `scaled_dot_product_attention` op. 
A new attribute `enable_gqa` has been added. As of now, only the
default value for the same is supported.

Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>

include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td

@@ -13734,7 +13734,7 @@ def Torch_AtenScaledDotProductAttentionOp : Torch_Op<"aten.scaled_dot_product_at
     HasValueSemantics,
     ReadOnly
   ]> {
-  let summary = "Generated op for `aten::scaled_dot_product_attention : (Tensor, Tensor, Tensor, Tensor?, float, bool, float?) -> (Tensor)`";
+  let summary = "Generated op for `aten::scaled_dot_product_attention : (Tensor, Tensor, Tensor, Tensor?, float, bool, float?, bool) -> (Tensor)`";
   let arguments = (ins
     AnyTorchTensorType:$query,
     AnyTorchTensorType:$key,
@@ -13742,7 +13742,8 @@ def Torch_AtenScaledDotProductAttentionOp : Torch_Op<"aten.scaled_dot_product_at
     AnyTorchOptionalTensorType:$attn_mask,
     Torch_FloatType:$dropout_p,
     Torch_BoolType:$is_causal,
-    AnyTorchOptionalFloatType:$scale
+    AnyTorchOptionalFloatType:$scale,
+    Torch_BoolType:$enable_gqa
   );
   let results = (outs
     AnyTorchOptionalTensorType:$result
@@ -13750,10 +13751,10 @@ def Torch_AtenScaledDotProductAttentionOp : Torch_Op<"aten.scaled_dot_product_at
   let hasCustomAssemblyFormat = 1;
   let extraClassDefinition = [{
     ParseResult AtenScaledDotProductAttentionOp::parse(OpAsmParser &parser, OperationState &result) {
-      return parseDefaultTorchOp(parser, result, 7, 1);
+      return parseDefaultTorchOp(parser, result, 8, 1);
     }
     void AtenScaledDotProductAttentionOp::print(OpAsmPrinter &printer) {
-      printDefaultTorchOp(printer, *this, 7, 1);
+      printDefaultTorchOp(printer, *this, 8, 1);
     }
   }];
 }

lib/Conversion/TorchToTMTensor/TorchToTMTensor.cpp

@@ -1582,6 +1582,7 @@ public:
     Value dropoutP = op.getDropoutP();
     Value isCausal = op.getIsCausal();
     Value scale = op.getScale();
+    Value enableGQA = op.getEnableGqa();
     Type elementType =
         cast<ShapedType>(adaptor.getQuery().getType()).getElementType();
 
@@ -1604,6 +1605,11 @@ public:
         return rewriter.notifyMatchFailure(op.getLoc(),
                                            "only default scale supported");
     }
+    bool isGQAEnabled;
+    if (!matchPattern(enableGQA, m_TorchConstantBool(&isGQAEnabled)) ||
+        isGQAEnabled)
+      return rewriter.notifyMatchFailure(
+          op.getLoc(), "grouped query attention not supported");
 
     auto opTy = cast<ValueTensorType>(op.getType()).toBuiltinTensor();
     auto query = adaptor.getQuery();

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -8832,7 +8832,7 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0 = call @__torch__.torch.jit._shape_functions.linear(%arg0, %arg1, %arg2) : (!torch.list<int>, !torch.list<int>, !torch.optional<list<int>>) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
 "  }\n"
-"  func.func @\"__torch_mlir_shape_fn.aten.scaled_dot_product_attention\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.list<int>, %arg3: !torch.optional<list<int>>, %arg4: !torch.float, %arg5: !torch.bool, %arg6: !torch.optional<float>) -> !torch.list<int> {\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.scaled_dot_product_attention\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.list<int>, %arg3: !torch.optional<list<int>>, %arg4: !torch.float, %arg5: !torch.bool, %arg6: !torch.optional<float>, %arg7: !torch.bool) -> !torch.list<int> {\n"
 "    %int-1 = torch.constant.int -1\n"
 "    %0 = torch.aten.__getitem__.t %arg2, %int-1 : !torch.list<int>, !torch.int -> !torch.int\n"
 "    %1 = torch.aten._set_item.t %arg0, %int-1, %0 : !torch.list<int>, !torch.int, !torch.int -> !torch.list<int>\n"
@@ -12446,7 +12446,7 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %int11 = torch.constant.int 11\n"
 "    return %int11 : !torch.int\n"
 "  }\n"
-"  func.func @\"__torch_mlir_dtype_fn.aten.scaled_dot_product_attention\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>, %arg2: !torch.tuple<int, int>, %arg3: !torch.optional<tuple<int, int>>, %arg4: !torch.float, %arg5: !torch.bool, %arg6: !torch.optional<float>) -> !torch.int {\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.scaled_dot_product_attention\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>, %arg2: !torch.tuple<int, int>, %arg3: !torch.optional<tuple<int, int>>, %arg4: !torch.float, %arg5: !torch.bool, %arg6: !torch.optional<float>, %arg7: !torch.bool) -> !torch.int {\n"
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    return %0#1 : !torch.int\n"
 "  }\n"

lib/Dialect/Torch/Transforms/ReduceOpVariants.cpp

@@ -246,6 +246,9 @@ void TorchMatchSpecializedBackendOp::populateSpecializedConversions(
           llvm::SmallVector<Value> newOperands{
               oldOperands[0], oldOperands[1], oldOperands[2], oldOperands[5],
               oldOperands[3], oldOperands[4], oldOperands[6]};
+          Value enableGQA =
+              rewriter.create<ConstantBoolOp>(op->getLoc(), false);
+          newOperands.push_back(enableGQA);
 
           auto newOp = rewriter.create<Torch::AtenScaledDotProductAttentionOp>(
               op.getLoc(), op->getResultTypes()[0], newOperands,

projects/pt1/e2e_testing/xfail_sets.py

@@ -33,6 +33,13 @@ LINALG_XFAIL_SET = COMMON_TORCH_MLIR_LOWERING_XFAILS | {
     "UnfoldModule_basic",
 }
 
+if torch_version_for_comparison() < version.parse("2.5.0.dev"):
+    LINALG_XFAIL_SET = LINALG_XFAIL_SET | {
+        # Error: 'torch.aten.scaled_dot_product_attention' op expected 8 operands, but found 7
+        "ScaledDotProductAttentionDifferentModule_basic",
+        "ScaledDotProductAttentionSameModule_basic",
+    }
+
 LINALG_CRASHING_SET = {
     # Runtime op verification: Out of bounds access
     "AtenDiagEmbedNegOffsetDiag_basic",

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py

@@ -1277,7 +1277,7 @@ def aten〇linear〡shape(input: List[int], weight: List[int], bias: Optional[Li
     Invocation(TensorOfShape(3, 2, 8, 4), TensorOfShape(3, 2, 8, 4), TensorOfShape(3, 2, 8, 4)), # Same shape
     Invocation(TensorOfShape(3, 2, 16, 8), TensorOfShape(3, 2, 8, 8), TensorOfShape(3, 2, 8, 4)), # Different shape
 ])
-def aten〇scaled_dot_product_attention〡shape(query: List[int], key: List[int], value: List[int], attn_mask: Optional[List[int]] = None, dropout_p: float = 0., is_causal: bool = False, scale: Optional[float] = None) -> List[int]:
+def aten〇scaled_dot_product_attention〡shape(query: List[int], key: List[int], value: List[int], attn_mask: Optional[List[int]] = None, dropout_p: float = 0., is_causal: bool = False, scale: Optional[float] = None, enable_gqa: bool = False) -> List[int]:
     outshape = query
     outshape[-1] = value[-1]
     return outshape
@@ -3558,7 +3558,7 @@ def aten〇isclose〡dtype(self_rank_dtype: Tuple[int, int], other_rank_dtype: T
     return torch.bool
 
 @check_dtype_function(_check_tensors_with_the_same_dtype(tensor_shapes=[(3, 4, 32, 16), (3, 4, 32, 16), (3, 4, 32, 16)]))
-def aten〇scaled_dot_product_attention〡dtype(query_rank_dtype: Tuple[int, int], key_rank_dtype: Tuple[int, int], value_rank_dtype: Tuple[int, int], attn_mask_rank_dtype: Optional[Tuple[int, int]] = None, dropout_p: float = 0., is_causal: bool = False, scale: Optional[float] = None) -> int:
+def aten〇scaled_dot_product_attention〡dtype(query_rank_dtype: Tuple[int, int], key_rank_dtype: Tuple[int, int], value_rank_dtype: Tuple[int, int], attn_mask_rank_dtype: Optional[Tuple[int, int]] = None, dropout_p: float = 0., is_causal: bool = False, scale: Optional[float] = None, enable_gqa: bool = False) -> int:
     _, query_dtype = query_rank_dtype
     return query_dtype
 

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py

@@ -988,7 +988,7 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     emit("aten::upsample_nearest2d : (Tensor, int[], float?, float?) -> (Tensor)")
     emit("aten::upsample_nearest2d.vec : (Tensor, int[]?, float[]?) -> (Tensor)")
     emit(
-        "aten::scaled_dot_product_attention : (Tensor, Tensor, Tensor, Tensor?, float, bool, float?) -> (Tensor)"
+        "aten::scaled_dot_product_attention : (Tensor, Tensor, Tensor, Tensor?, float, bool, float?, bool) -> (Tensor)"
     )
     emit("aten::grid_sampler : (Tensor, Tensor, int, int, bool) -> (Tensor)")
 

pytorch-hash.txt

@@ -1 +1 @@
-d6ea1eb2bc8ba770fd5a689a30e234837df27384
+748db193d71a1c29471a87c7841da6a5a0a0dbae

pytorch-requirements.txt

@@ -1,3 +1,3 @@
 -f https://download.pytorch.org/whl/nightly/cpu/torch/
 --pre
-torch==2.5.0.dev20240804
+torch==2.5.0.dev20240818

test/Dialect/Torch/reduce-op-variants.mlir

@@ -193,7 +193,7 @@ func.func @torch.aten.bernoulli_.float(%t: !torch.tensor) -> !torch.tensor {
 // CHECK:      %[[FALSE:.+]] = torch.constant.bool false
 // CHECK:      %[[NONE0:.+]] = torch.constant.none
 // CHECK:      %[[NONE1:.+]] = torch.constant.none
-// CHECK:      %[[ATTEN:.+]] = torch.aten.scaled_dot_product_attention %[[ARG0]], %[[ARG1]], %[[ARG2]], %[[NONE0]], %[[ZERO]], %[[FALSE]], %[[NONE1]]
+// CHECK:      %[[ATTEN:.+]] = torch.aten.scaled_dot_product_attention %[[ARG0]], %[[ARG1]], %[[ARG2]], %[[NONE0]], %[[ZERO]], %[[FALSE]], %[[NONE1]], %[[FALSE]]
 // CHECK:      return %[[ATTEN]]
 func.func @scaled_dot_product_flash_attention_for_cpu(%arg0: !torch.vtensor<[1,1,5,5],f32>, %arg1: !torch.vtensor<[1,1,5,5],f32>, %arg2: !torch.vtensor<[1,1,5,5],f32>) -> !torch.vtensor<[1,1,5,5],f32> {
   %float0.000000e00 = torch.constant.float 0.000000e+00

torchvision-requirements.txt

@@ -1,3 +1,3 @@
 -f https://download.pytorch.org/whl/nightly/cpu/torchvision/
 --pre
-torchvision==0.20.0.dev20240804
+torchvision==0.20.0.dev20240818