ROCm/rocm-libraries

Summary Expose the remaining hipDNN Graph lifecycle, plan-management, inspection, and serialization APIs in the Python bindings. Parent: #10579 Required API Add Python bindings for: build checkTensorUidsSet getTensorsByUid getTensorsByName set_override_shape_enabled is_override_shape_enabled The execution overload accepting runtime shape and stride overrides. to_compiled_plan_binary from_compiled_plan_binary get_execution_plan_count execute_plan_at_index get_workspace_size_plan_at_index get_plan_name_at_index build_plan_at_index get_autotune_workspace_size get_plan_name Use Python-friendly return values for C++ output parameters and preserve optional values and error reporting. The override-shape methods must remain consistent with the existing HIPDNN_ENABLE_SDPA feature guard. Tests Add focused tests for: build and the complete standard graph lifecycle. Tensor UID/name inspection. Plan count, plan selection, plan execution, workspace, and plan-name queries. Compiled-plan serializat

Summary Complete the Python bindings for hipDNN TensorAttributes and scalar tensor metadata. The current binding covers basic dimensions, strides, dtype, UID, name, virtual status, and a float-only set_value. It does not expose the complete public tensor metadata and pass-by-value API needed for full frontend parity or DLPack integration. Parent: #10579 Required API Expose the remaining public tensor metadata methods, including: Scalar tensor construction for all supported value types. get_is_pass_by_value get_pass_by_value get_compile_time_constant get_has_compile_time_constant get_is_runtime_pass_by_value set_as_runtime_parameter set_is_pass_by_value set_compile_time_constant get_value_variant clear_value Ragged-offset getters, setters, and presence checks. Alignment getters and setters. Any remaining public data-type, dimension, stride, output, virtual, UID, and name accessors. Use Python-friendly conversions for C++ scalar variants and optionals. Preserve existing Tensor.create()

Suggestion Description As part of enabling relaxed affine tensor shape requirements for BatchNorm in this PR in hipDNN, the corresponding frontend changes have currently been implemented only for the BatchNorm operation. To fully support this feature across all BatchNorm operations, the same functionality also needs to be implemented in: BatchnormBackwardNode, BatchnormInferenceNode and BatchnormInferenceNodeVarianceExt. Without these changes, users cannot take full advantage of the relaxed affine tensor shape requirements for the batchnorm operation.

While investigating #8530, we reviewed other hipDNN frontend node property inference paths for cases where missing output tensor metadata may be ambiguous or inconsistent with cuDNN frontend behavior. SDPA backward is not ambiguous in the same way as convolution backward data/weight, but it still has validation and layout-policy gaps that should be tracked. Current behavior SdpaBwdNode::infer_properties_node() infers missing gradient output metadata as: dQ.dim = Q.dim dK.dim = K.dim dV.dim = V.dim missing dQ/dK/dV strides are generated with the default packed row-major/BHSD policy This shape inference is deterministic and mathematically correct for SDPA backward, including GQA/MQA: the gradients must match the original tensors they differentiate. This differs from cuDNN frontend, which generally expects backward output descriptors to be supplied, but it is not inherently unsafe the way convolution dx / dw inference can be unsafe. The gap is that hipDNN currently accepts caller-suppli

Context From PR #6296 review by @mousdahl-amd — merged as-is with this tracked as a follow-up. Problem The integration round-trip tests (JSON and binary serialize/deserialize) manually compare every field of the deserialized graph against expected values. This is verbose and fragile — each new field requires updating every round-trip test. Proposal Implement Equals, (Possibly call it LogicallyEquals) for: Graph Node (and its concrete subtypes or via virtual dispatch) Tensor operator == likely should also compare the backend descriptors, and for the checks we want here, it should be skipped. Likely want to split into two things with overlapping support. LogicallyEquals which compares logical equivalency of the graph state (Graph properties + Nodes + Tensor definitions all match). Operator==, that checks LogicallyEquals && the descriptor pointers / lowering state. Tests can use LogicallyEquals, and consumers can use LogicallyEquals for checking if graphs represent the same operatio