[xnn update prep] deprecate sdpa

ghstack-source-id: 929a8bc
ghstack-comment-id: 2957149191
Pull Request resolved: #11506

Author: mcr229

backends/xnnpack/operators/__init__.py

@@ -40,7 +40,6 @@
     op_quant_dequant,
     op_relu,
     op_rsqrt,
-    op_sdpa,
     op_sigmoid,
     op_skip_ops,
     op_slice_copy,

backends/xnnpack/operators/op_sdpa.py

@@ -44,7 +44,6 @@
     QuantizedPerTensorConfig,
     ReciprocalSquareRootConfig,
     ReLUConfig,
-    # SDPAConfig, TODO: D60553559: preserving SDPA for fairseq fails
     SigmoidConfig,
     SliceCopyConfig,
     SoftmaxConfig,
@@ -103,7 +102,6 @@
     ReciprocalSquareRootConfig,
     ReLUConfig,
     TanhConfig,
-    # SDPAConfig, TODO: D60553559: preserving SDPA for fairseq fails
     SigmoidConfig,
     SliceCopyConfig,
     SoftmaxConfig,

backends/xnnpack/partition/config/__init__.py

@@ -564,33 +564,3 @@ class BMMConfig(GenericNodePartitionerConfig):
 
     def supported_precision_types(self) -> List[ConfigPrecisionType]:
         return [ConfigPrecisionType.FP32]
-
-
-class SDPAConfig(GenericNodePartitionerConfig):
-    target_name = "scaled_dot_product_attention.default"
-
-    def check_constraints(self, node: torch.fx.Node, ep: ExportedProgram) -> bool:
-        """
-        Requires Mask to have Rank 2
-        """
-        if not self.check_common_constraints(node, ep):
-            return False
-
-        if len(node.all_input_nodes) < 4:
-            return False
-        mask_node = node.all_input_nodes[3]
-        mask_rank = mask_node.meta["val"].dim()
-        if mask_rank != 2:
-            why(
-                node,
-                reason=f"mask must have rank 2, got mask of rank {mask_rank}",
-            )
-            return False
-
-        return True
-
-    def get_original_aten(self) -> Optional[torch._ops.OpOverload]:
-        return torch.ops.aten.scaled_dot_product_attention.default
-
-    def supported_precision_types(self) -> List[ConfigPrecisionType]:
-        return [ConfigPrecisionType.FP32]

backends/xnnpack/partition/config/generic_node_configs.py

@@ -1423,42 +1423,6 @@ Error defineStaticSliceNode(
   return Error::Ok;
 }
 
-/*
-Defines Scaled Dot Product Attention (SDPA) node into the subgraph,
-using the remapped ids to map the serialized ids,
-to the new ids generated when defining the tensor value
-*/
-Error defineScaledDotProductAttentionNode(
-    xnn_subgraph_t subgraph_ptr,
-    const std::unordered_map<uint32_t, uint32_t>& remapped_ids,
-    const NodePtr node,
-    const fb_xnnpack::XNNGraph* graph) noexcept {
-  MAYBE_UNUSED(graph);
-
-  auto graph_node = node->xnode_union_as_XNNScaledDotProductAttention();
-
-  xnn_status status = xnn_define_scaled_dot_product_attention(
-      subgraph_ptr,
-      xnn_attention_logits_cap_type_none, // cap_type
-      nullptr, // cap_value - not used
-      remapped_ids.at(graph_node->query_id()),
-      remapped_ids.at(graph_node->key_id()),
-      remapped_ids.at(graph_node->value_id()),
-      remapped_ids.at(graph_node->scale_id()),
-      remapped_ids.at(graph_node->mask_id()),
-      remapped_ids.at(graph_node->output_id()),
-      graph_node->flags());
-
-  ET_CHECK_OR_RETURN_ERROR(
-      status == xnn_status_success,
-      Internal,
-      "Failed to create SDPA node %i with code: %s",
-      node->debug_handle(),
-      xnn_status_to_string(status));
-
-  return Error::Ok;
-}
-
 /*
 Defines batch matrix multiply node into the subgraph,
 using the remapped ids to map the serialized ids,
@@ -1788,7 +1752,6 @@ DefineNodeFunc getDefineNodeFunc(fb_xnnpack::XNodeUnion nodeType) {
     _DEFINE(Concatenate4)
     _DEFINE(Concatenate5)
     _DEFINE(StaticSlice)
-    _DEFINE(ScaledDotProductAttention)
     _DEFINE(BatchMatrixMultiply)
     case fb_xnnpack::XNodeUnion::NONE:
     default: // Adding here as a catch all, just in case