onnxruntime/core/session/inference_session.cc (14 lines): - line 48: #ifdef USE_DML // TODO: This is necessary for the workaround in TransformGraph - line 83: // TODO: use LoggingManager::GetTimestamp and date::operator<< - line 116: if (is_verbose) { // TODO: should we disable this if the number of nodes is above a certain threshold? - line 438: // TODO: Currently we have no way to transport this error to the API user - line 584: // TODO: Investigate whether we should set a max size, and clarify the cost of having a buffer smaller than - line 607: // TODO: Do we need to catch any std::exceptions from creating/writing to disk and convert to Status codes? - line 891: // TODO: this is a temporary workaround to apply the DML EP's custom graph transformer prior to partitioning. This - line 1305: // TODO: we could refactor the allocators to not require the call to GetAllocator but that change is much bigger - line 1307: // TODO (contd.) We could also possibly absorb the per-thread logic in a new allocator decorator that derives - line 1704: // TODO add more validation here like checking shape of the allocated buffers - line 1736: // TODO: only call OnRunStart for all providers in-use - line 1849: // TODO: only call OnRunStart for all providers in-use - line 2012: // TODO should Run() call io_binding.SynchronizeInputs() or should it let the callers do it? - line 2137: // TODO add other post load processing here onnxruntime/python/onnxruntime_pybind_state.cc (11 lines): - line 61: // TODO: we may delay-init this variable - line 86: // TODO: How to handle unload failure ? - line 909: // TODO remove deprecated global config - line 914: // TODO remove deprecated global config - line 933: // TODO remove deprecated global config - line 940: // TODO remove deprecated global config - line 957: // TODO remove deprecated global config - line 962: // TODO remove deprecated global config - line 972: // TODO remove deprecated global config - line 983: // TODO remove deprecated global config - line 990: // TODO remove deprecated global config onnxruntime/python/onnxruntime_pybind_state_common.cc (10 lines): - line 13: // TODO remove deprecated global config - line 18: // TODO remove deprecated global config - line 22: // TODO remove deprecated global config - line 24: // TODO remove deprecated global config - line 28: // TODO remove deprecated global config - line 30: // TODO remove deprecated global config - line 33: // TODO remove deprecated global config - line 38: // TODO remove deprecated global config - line 40: // TODO remove deprecated global config - line 43: // TODO remove deprecated global config onnxruntime/core/providers/nuphar/nuphar_execution_provider.cc (10 lines): - line 6: #include "core/codegen/passes/utils/ort_tvm_utils.h" // TODO remove this after removing tvm::runtime - line 10: #include "core/providers/nuphar/common/analysis/shape_expr.h" // TODO: remove this shape_expr after shape_infernece refinement - line 19: #include // TODO remove this after removing tvm::runtime - line 72: codegen_target_ = std::make_unique(target_str, 128, 1); // TODO: use real values - line 115: // TODO add multi-target support - line 134: // TODO: remove - line 135: handle->allow_unaligned_buffers = info.allow_unaligned_buffers; // TODO remove this - line 218: // TODO: enable support for Indices - line 222: // TODO: enable support for non-default dilations - line 432: // TODO move compilation part out of create_state_func to above onnxruntime/python/onnxruntime_pybind_state_common.h (8 lines): - line 141: // TODO remove deprecated global config - line 151: // TODO remove deprecated global config - line 179: // TODO remove deprecated global config - line 181: // TODO remove deprecated global config - line 194: // TODO remove deprecated global config - line 196: // TODO remove deprecated global config - line 217: // TODO remove deprecated global config - line 219: // TODO remove deprecated global config orttraining/orttraining/eager/ort_aten.cpp (8 lines): - line 113: // TODO: support more types - line 330: // TODO: validate options and memory format - line 331: // TODO: figure out how to get the correct element type. - line 357: // TODO: handle stride - line 358: // TODO: how to handle type conversion - line 361: // TODO: how to support layout - line 382: // TODO: support stride - line 493: // TODO: enhance opgen.py to support inplace binary operations. js/web/lib/onnxjs/backends/webgl/glsl-coordinate-lib.ts (8 lines): - line 28: // TODO return these only when packing is enabled. - line 112: // TODO we need this to properly return a packed vec4 from kernels. - line 163: // TODO we need this to properly return a packed vec4 from kernels. - line 540: // TODO: swap texR, texC order in following function so row is corresponding to u and column is corresponding to - line 788: // TODO support more dimensionalities - line 996: // TODO: modify row/col order for other dimensions. - line 1084: // TODO: revisit the logic here to make it simpler - line 1120: // TODO: re-enable this shortcut once the index calculation bug is fixed. orttraining/orttraining/core/graph/mixed_precision_transformer.cc (7 lines): - line 357: // TODO: Why do we need to handle "Cast" here? - line 401: // TODO: Shouldn't we check stage2_fp32_node_args to conditionally transform this? - line 407: // TODO: This currently assumes that all outputs of FP32 ops are FP32. - line 430: // TODO: Ideally, we should not need to transform a function-body here. - line 447: // TODO: eliminate use of const_casts - line 449: // TODO: The resolve below is likely unnecessary. - line 464: // TODO: fix const_cast; Graph doesn't provide us a method "GetMutableInputs". onnxruntime/core/framework/allocation_planner.cc (7 lines): - line 364: // TODO: This should probably be defined to be the equality operator on TensorShapeProto. - line 411: /* TODO: we can improve this if the concrete shapes are known for both as below. - line 448: // TODO this should be an error case, needs more investigation - line 714: // TODO: Identify error cases where-in an initializer is used on different - line 756: // TODO: Move away from usage of vector of `OrtMemoryInfo`s per weight (initializer) - line 1112: //TODO: should be size_t - line 1146: // TODO: unclear why we should go through a string-representation of type orttraining/orttraining/core/graph/training_op_defs.cc (7 lines): - line 383: // TODO: This is copied from onnx schemas. When the change is in and we update this can be removed. - line 874: // TODO: Move this to the right location. Its only here for quick experimentation. - line 875: // TODO: Use the mutli weight / grad version. - line 893: // TODO: This is copied from onnx schemas. When the change is in and we update this can be removed. - line 1095: // TODO: Depreacate this schema when training support is udpated to opset-12 - line 1192: // TODO: push this to ONNX - line 1723: TODO: Note that the above doesn't handle the case where training_mode=false and a non-zero onnxruntime/core/providers/nuphar/compiler/nuphar_codegen_ctx.cc (7 lines): - line 8: #include "core/codegen/mti/mti_tvm_utils.h" // TODO: remove this after decoupling layout compile and run - line 11: #include "core/codegen/passes/utils/ort_tvm_utils.h" // TODO: remove this after decoupling layout compile and run - line 12: #include // TODO: remove this after decoupling layout compile and run - line 46: // TODO: remove this. - line 77: // TODO: remove this. - line 160: // TODO: change to delayed call - line 163: // TODO: change to delayed call js/web/lib/onnxjs/backends/webgl/webgl-context.ts (6 lines): - line 145: // TODO: Check if framebuffer is ready - line 153: // TODO: Implement logic to check if the framebuffer is ready - line 508: // TODO: add webgl 1 handling. - line 520: // TODO: add webgl 1 handling. - line 534: // TODO: add webgl 1 handling. - line 548: // TODO: add webgl 1 handling. orttraining/orttraining/core/session/training_session.cc (6 lines): - line 87: // TODO make OptimizerNodeConfig::loss_scale_input_name optional - line 112: // TODO make OptimizerGraphConfig::loss_scale_input_name optional - line 481: // TODO: this is a temp workaround for removing rank tensor before adding optimizer. - line 716: // TODO: ideally we can just reuse the CPU EP registered with the session, but in the training session case - line 964: // TODO: support more types than float - line 1006: std::remove(ToUTF8String(model_uri).c_str()); // TODO would be good to have something like RemoveFile(PathString) orttraining/orttraining/eager/opgen/opgen/generator.py (6 lines): - line 260: # FIXME: find the first at::Tensor param anywhere in the signature - line 310: # FIXME: warn if we have not consumed all torch parameters (either as - line 393: # TODO: This assumes that the first output corresponds to the first input. - line 433: # TODO: Pick the right "out" Torch parameter; do not assume the first one - line 434: # TODO: Handle mutliple results - line 435: # TODO: Assert return type onnxruntime/core/providers/cpu/tensor/tile.cc (5 lines): - line 174: // TODO: Handle string copies when the kernel eventually supports string type. - line 192: // TODO: Handle string copies when the kernel eventually supports string type. - line 202: // TODO: Add multi-threading logic if num_of_copies_per_batch is large enough - line 211: // TODO: Multi-thread if needed - line 262: // TODO: Support 'string' and 'float16' types for completeness onnxruntime/contrib_ops/cpu/transformers/gpt_subgraph.cc (5 lines): - line 25: // Could reduce the chance of arithmetic overflow. TODO: fix it - line 118: // TODO: support float16 - line 272: // TODO: Try expand inputs/outputs after first subgraph call instead. That may get better peroformance, but more complex to implement. - line 336: // TODO: support float16 - line 391: // TODO: Reuse buffer for input_ids and position_ids to reduce memory allocation. onnxruntime/core/platform/windows/env.cc (5 lines): - line 101: // TODO: should I try to use SetThreadSelectedCpuSets? - line 542: // TODO: in most cases, the path name is a relative path and the behavior of the following line of code is undefined. - line 560: // TODO: trim the ending '\r' and/or '\n' - line 593: // TODO: trim the ending '\r' and/or '\n' - line 635: // TODO: Understand the reason for failure by calling GetLastError(). onnxruntime/core/providers/nuphar/scripts/model_editor.py (5 lines): - line 59: # TODO: support these activation attributes - line 216: # TODO: will input order get messed up - line 268: # TODO: support peephole - line 276: # TODO: implement input_forget = 1 - line 700: # TODO: assert the the initializer raw data is not 0 (False) onnxruntime/core/graph/function.cc (5 lines): - line 13: // TODO: Add this to onnx instead of adding it here. - line 146: // TODO: revisit to see if we can eliminate typeconstraint step - line 432: //TODO: for better performance, we could try to transfer the nodes in parent graph to sub-graph directly, - line 470: //TODO: if we reuse the nodes in parent graph, maybe we don't need to resolve it. - line 686: // TODO: Need a proper scheme to generate unique names to avoid name-collision. orttraining/orttraining/python/training/ortmodule/_graph_execution_manager.py (4 lines): - line 157: # TODO: remove after PyTorch ONNX exporter supports VAR_KEYWORD parameters. - line 337: TODO: How to support dynamic axes? Dimensions are determined by samples - line 430: # TODO: PyTorch exporter bug: changes the initializer order in ONNX model - line 453: # TODO: Explore ways to make self._graph_info.initializer_names and self._graph_info.initializer_names_to_train orttraining/orttraining/core/graph/pipeline_transformer.cc (4 lines): - line 857: // TODO: re-visit the different cases to see if there are other situations aside from constant ops. - line 1329: // TODO: Should we consider weights here too? - line 1363: // TODO: test case in which a tensor is produced by a forward op, stashed - line 1757: // TODO: We should consider checking the cut is valid --- Cut Infos should onnxruntime/core/providers/nuphar/compiler/x86/scheduler/tensorize_schedule.cc (4 lines): - line 321: // TODO: add 16bit tensorize GEMV/GEMM for AVX512 - line 323: // TODO: add 16bit tensorize GEMM for AVX2 - line 325: // TODO: release 16bit tensorize GEMV for AVX2 - line 354: // TODO: enable 16 bit tensorization winml/lib/Api.Image/VideoFrameToTensorConverter.cpp (4 lines): - line 214: // TODO: Scale during the tensorization phase instead of using the video frame pipeline when the input bounds are not the same size as the tensor - line 280: // TODO: merge all videoframes to a single DX12Texture Resource before call ConvertDX12TextureToGPUTensor. - line 481: // TODO: Scale during the tensorization phase instead of using the video frame pipeline when the input bounds are not the same size as the tensor - line 509: // TODO: Make an allocator for upload heaps onnxruntime/core/optimizer/propagate_cast_ops.cc (4 lines): - line 568: // TODO: If the specified optimization is greater than 1 then insert a Cast to the - line 633: // TODO: If the specified level of the optimization is greater than 1 then - line 753: // TODO need a heuristic when to insert FP32 Cast - line 953: // TODO Here we require the all floating point inputs are generated by an immediate orttraining/orttraining/training_ops/cuda/loss/softmaxcrossentropy_impl.h (4 lines): - line 71: // TODO: implement reduction type of NONE - line 82: // TODO: implement reduction type of NONE - line 93: // TODO: implement reduction type of NONE - line 104: // TODO: implement reduction type of NONE orttraining/orttraining/python/orttraining_pybind_state.cc (4 lines): - line 129: // TODO: this method does not handle parallel optimization. - line 132: //TODO tix, refactor the mpi related code to populate all fields correctly by default. - line 249: // TODO: this mapping is temporary. - line 258: // TODO reduction types orttraining/orttraining/python/training/orttrainer.py (4 lines): - line 151: # TODO: Remove when experimental checkpoint functions are removed. - line 168: # TODO: Move this to model_desc_validation.py - line 203: # TODO: Remove when experimental checkpoint functions are removed. - line 790: # TODO: Remove when experimental checkpoint functions are removed onnxruntime/core/providers/nuphar/compiler/nuphar_handle.h (4 lines): - line 10: #include "core/framework/allocator.h" // TODO: get rid of this - line 11: #include "core/providers/nuphar/compiler/traverse_shape_infer.h" // TODO: get rid of this - line 24: // TODO: make this the Base - line 25: // TODO: create one for nuphar onnxruntime/core/graph/graph.cc (4 lines): - line 467: // TODO: What is the plan for optional sequence tensors ? - line 2570: // TODO: Handle optional attribute but no default value specified in op definition. - line 3818: // TODO: Does this need (and maybe AllocateNode) to be threadsafe so nodes_ and num_of_nodes_ managed more carefully? - line 4230: // TODO: We could decide that an ORT model is load only even in a full build, onnxruntime/core/providers/nuphar/mti_x86/math/matmul_ops.cc (4 lines): - line 128: nullptr); // TODO: use thread pool from OpContext - line 180: nullptr); // TODO: use thread pool from OpContext - line 220: // TODO: add support for mixed precisions - line 260: // TODO: add support for mixed precisions orttraining/orttraining/python/ort_trainer.py (3 lines): - line 86: # TODO: get the constant string via pybind. - line 259: # TODO: to unpack the input argument. - line 645: # TODO: accept loss_fn as an onnx model. build self.onnx_model_ with model and loss_fn orttraining/orttraining/training_ops/cuda/optimizer/lamb.cc (3 lines): - line 48: // TODO: Once Schema is checked in to onnx lets fix this to match that - line 147: // TODO: temporary hack until View is improved (it doesn't work with Alias) - line 624: // TODO: temporary hack until View is improved (it doesn't work with Alias) onnxruntime/core/providers/nuphar/compiler/initializer_info.h (3 lines): - line 8: // TODO: move to nuphar - line 12: // TODO: move it to weight layout place - line 14: const Tensor* marshalled_initializer = nullptr; // TODO: change it to unique_ptr include/onnxruntime/core/session/onnxruntime_c_api.h (3 lines): - line 658: // TODO: document the path separator convention? '/' vs '\' - line 659: // TODO: should specify the access characteristics of model_path. Is this read only during the - line 3318: // TODO: Support 'Variadic' inputs/outputs onnxruntime/core/providers/nuphar/common/analysis/output_alias_analysis.cc (3 lines): - line 20: // TODO: change identity to other alias - line 32: // TODO: remove once we have a better solution for alias optimization - line 46: // TODO: please reimplement output alias using the right algorithm. onnxruntime/contrib_ops/cpu/transformers/logits_processor.cc (3 lines): - line 5: // Could reduce the chance of arithmetic overflow. TODO: fix it - line 107: // TODO: build N-Gram index (hash table with prefix of length NGram - 1 as key, and list of last word of NGram as value) for fast matching. - line 135: // vocab_mask shape (vocab_size). TODO: support shape (batch_size, vocab_size) onnxruntime/python/tools/transformers/onnx_model.py (3 lines): - line 522: # TODO: import from onnxconverter_common when it is stable - line 755: # TODO: handle inner graph - line 862: #TODO: support graph_topological_sort() in subgraphs onnxruntime/python/tools/transformers/longformer/convert_longformer_to_onnx.py (3 lines): - line 46: # TODO: search the directory to find correct output filename of "python setup.py install" when python version is not 3.6 - line 125: # TODO: add postprocess of ONNX model to calculate based on graph input: input_mask = (attention_mask - 1) * 10000.0 - line 126: # TODO: add postprocess of ONNX model to use graph input directly: glboal_mask = global_attention_mask setup.py (3 lines): - line 57: # TODO: how to support multiple CUDA versions? - line 276: libs = ['onnxruntime_pybind11_state.so', 'libdnnl.2.dylib', 'mimalloc.so'] # TODO add libmklml and libiomp5 later. - line 496: # TODO: this is the last time we have to do this!!! csharp/src/Microsoft.ML.OnnxRuntime/Tensors/Tensor.shared.cs (3 lines): - line 451: // TODO: allow specification of axis1 and axis2? - line 827: // TODO: allow specification of axis1 and axis2? - line 940: // TODO: allow specification of axis1 and axis2? onnxruntime/core/providers/cuda/fpgeneric.cu (3 lines): - line 11: // NV_TODO: optimize speed -- pass things needed in, optimize kernel speed, add half2 - line 12: // NV_TODO: investigate cub support for half - line 25: // TODO - refactor the function with similar logic in Transpose3DKernel using 16x16 Tile js/web/lib/onnxjs/backends/webgl/ops/conv.ts (3 lines): - line 119: // TODO : Make this generic enough to compute default attributes for multi-dimensional conv - line 137: // TODO : Need to add support for multi-dimensional conv - line 176: // TODO : Need to add support for float64 onnxruntime/core/providers/nuphar/compiler/x86/op_ir_creator/math/quantize/matmul_integer.cc (3 lines): - line 16: #include "core/common/cpuid_info.h" // TODO: refactor to control through config - line 138: //TODO: change to use MLAS when no layout could apply - line 256: //TODO: change to use MLAS when no layout could apply onnxruntime/python/onnxruntime_pybind_ortvalue.cc (3 lines): - line 52: // TODO: Add check to ensure that string arrays are not passed - we currently don't support string tensors in CUDA - line 61: // TODO: Add check to ensure that string arrays are not passed - we currently don't support string tensors in CUDA - line 129: // TODO: Assumes that the OrtValue is a Tensor, make this generic to handle non-Tensors include/onnxruntime/core/optimizer/rule_based_graph_transformer.h (3 lines): - line 24: @TODO: Is a bottom-up traversal more efficient? - line 25: @TODO: Is it worth adding the max number of passes a rule should be applied for? - line 26: @TODO: We need to define a contract about whether a rewrite rule is allowed to leave onnxruntime/core/framework/session_state.cc (3 lines): - line 348: // TODO: Check if some version of the ONNX IR allows op_type to be empty - line 1116: // TODO: Re-visit this method if more subgraph ops get accepted into ONNX - line 1336: // TODO: make it better tools/python/util/ort_format_model/operator_type_usage_processors.py (3 lines): - line 370: # TODO - review and add ML ops as needed - line 417: # TODO: Only apply required types to the global type list and ignore if it's model based per-op type reduction - line 455: TODO: Currently the type tracking is not specific to a version of the operator. js/web/lib/onnxjs/backends/webgl/op-resolve-rules.ts (3 lines): - line 46: // TODO: support new attributes for AveragePool-10 - line 78: // TODO: support new attributes for MaxPool-8 and MaxPool-10 - line 102: ['Slice', '', '10+', sliceV10], // TODO: support 'steps' for Slice-10 winml/lib/Api.Experimental/LearningModelOperator.cpp (3 lines): - line 37: // TODO Validate against allowed operator input NAMES. The types are not deduced. - line 44: // TODO Validate against allowed operator input NAMES. The types are not deduced. - line 53: // TODO Validate against allowed operator output NAMES. The types are not deduced. onnxruntime/core/graph/contrib_ops/contrib_defs.cc (3 lines): - line 145: // TODO: throw exception? - line 517: // TODO does the extra output need any changes? - line 2202: // TODO: replace with ONNX one when that one is fixed onnxruntime/core/framework/kernel_def_builder.cc (3 lines): - line 77: // TODO: Tell user why it has conflicts - line 78: // TODO: Investigate why IsConflict() was not triggered when there were duplicate Tile CUDA - line 244: // TODO: validate inputs. onnxruntime/core/providers/nuphar/nuphar_execution_provider.h (3 lines): - line 129: // TODO move this to another place - line 131: // TODO: move all tvm related code a manager - line 147: // TODO: support multi-target support onnxruntime/contrib_ops/cpu/transformers/beam_search.cc (3 lines): - line 33: // Could reduce the chance of arithmetic overflow. TODO: fix it - line 405: // TODO: use thread pool to parrellel - line 573: beam_scorer_->Initialize(allocator_, parameters_->sequence_length); // TODO: use temp_space_allocator orttraining/orttraining/python/training/ortmodule/_utils.py (3 lines): - line 30: # TODO: Current DLPack doesn't support bool and PyTorch disables converting bool tensor to DLPack in recent commit. - line 57: # TODO: Current DLPack doesn't support bool and PyTorch disables converting bool tensor to DLPack in recent commit. - line 261: # TODO: Rework by implementing the "__getattribute__" method. onnxruntime/core/providers/nuphar/partition/subgraph_partitioner.cc (3 lines): - line 21: // TODO remove this after rewriting Cost function and ForcePartition - line 27: // TODO: change the Cost and ForcePartition to a more complex form. - line 31: // TODO replace NodeUseCount approximation onnxruntime/core/providers/cuda/cuda_execution_provider.cc (3 lines): - line 2148: // TODO: Check if we can adopt a similar approach to deal with asymmetric pads in 'ConvTranspose' - line 2178: // TODO: Remove this after we have supported asymmetric padding in the CUDA ConvTranspose kernel - line 2327: // TODO: this is actually used for the cuda kernels which explicitly ask for inputs from CPU. orttraining/orttraining/core/framework/checkpointing.cc (3 lines): - line 72: // TODO is the encoding correct? https://github.com/onnx/onnx/issues/2392 - line 81: // TODO need to ensure the data is written in little-endian format... - line 247: // TODO is the encoding correct? https://github.com/onnx/onnx/issues/2392 tools/ci_build/build.py (3 lines): - line 847: # TODO: this variable is not really needed - line 1451: # TODO: use run_orttraining_test_orttrainer_frontend_separately to work around a sporadic segfault. - line 1479: # TODO: this test will be replaced with convergence test ported from backend onnxruntime/core/session/onnxruntime_c_api.cc (3 lines): - line 617: // TODO: Could move to a helper - line 1967: // TODO: there is no need to manually malloc/free these memory, it is insecure - line 1994: // TODO: we don't really need the second parameter onnxruntime/core/optimizer/transformer_memcpy.cc (3 lines): - line 74: // TODO: We probably need to do the recursion inline when processing the main graph in order to maximize efficiency. - line 241: // TODO: copy between devices? i.e. multiple GPUs - line 350: // TODO: The copy being made is possibly redundant if this occurs in a subgraph onnxruntime/core/providers/nnapi/nnapi_builtin/builders/op_builder.cc (3 lines): - line 260: // TODO support other data types - line 345: // TODO support other data types - line 880: // TODO add quantized matmul when reshape support quantized input onnxruntime/core/optimizer/insert_cast_transformer.cc (3 lines): - line 22: // TODO: a better check is to check does the CPU kernel with float exist or not. - line 405: // TODO 1: Check if the kernel available - line 406: // TODO 2: There is an inherent assumption that if we cast a cpu op's input from float16 to float onnxruntime/core/framework/graph_partitioner.cc (3 lines): - line 192: // TODO: when the graph contain a function node, and user pass in the dll which could - line 311: // TODO: Graph::FuseSubGraph and Graph::FinalizeFuseSubGraph should now create valid edges so this call to - line 443: // TODO: Could avoid the topological sort in the GraphViewer ctor by constructing from an existing cmake/onnxruntime_unittests.cmake (3 lines): - line 38: #TODO: fix the warnings, they are dangerous - line 289: # TODO: Add anything that is needed for testing a minimal build - line 618: #TODO: fix the warnings, they are dangerous onnxruntime/core/codegen/passes/op_ir_creator/nn/conv.cc (3 lines): - line 50: // TODO: handle symbolic dim - line 99: // FIXME: This will trigger a llvm buffer overflow when group is too large - line 100: // TODO: fix this change it to batched gemm/conv onnxruntime/core/framework/tensorprotoutils.cc (3 lines): - line 263: // TODO: complex64 complex128 - line 638: // TODO The line above has const-correctness issues. Below is a possible fix which copies the tensor_proto data - line 699: // TODO: Change the current interface to take Path object for model path js/common/lib/inference-session.ts (2 lines): - line 186: // TODO: add flags - line 190: // TODO: add flags onnxruntime/python/tools/transformers/longformer/benchmark_longformer.py (2 lines): - line 65: "engine": "torch", #TODO: test torchscript - line 234: # TODO: call convert_longformer_to_onnx to export onnx instead. onnxruntime/core/framework/data_types.cc (2 lines): - line 317: //TODO: Fix the warning - line 866: // TODO: name() method of `type_info` class is implementation dependent js/node/lib/backend.ts (2 lines): - line 29: // TODO: implement profiling - line 32: // TODO: implement profiling orttraining/orttraining/python/training/ortmodule/torch_cpp_extensions/cuda/fused_ops/multi_tensor_apply.cuh (2 lines): - line 20: // TODO: Kernel arg size limit may be <4KB for some other cards (ie Jetson) - line 61: // TODO: Print which tensor fails. js/node/src/session_options_helper.cc (2 lines): - line 38: // TODO: handling CPU EP options - line 40: // TODO: handling Cuda EP options js/web/lib/wasm/session-handler.ts (2 lines): - line 65: // TODO: support pre-allocated output - line 84: // TODO: implement profiling cmake/onnxruntime_framework.cmake (2 lines): - line 43: # TODO: for now, core framework depends on CUDA. It should be moved to TensorRT EP - line 44: # TODO: provider_bridge_ort.cc should not include nccl.h include/onnxruntime/core/framework/execution_provider.h (2 lines): - line 179: // TODO: temparary sulotion, need to unify the interface in EP and AllocatorManager - line 268: TODO: Used for CUDA & TRT only for now, will have one more PR to apply this for all EPs. js/web/lib/onnxjs/backends/webgl/ops/concat.ts (2 lines): - line 117: // TODO: Implement BinarySearch in GLSL - line 179: // TODO: Support string concat js/node/src/tensor_helper.cc (2 lines): - line 187: // TODO: change to TypedArray::ByteLength() in next node-addon-api release. - line 250: // TODO: optimize memory winml/lib/Api/ImageFeatureValue.cpp (2 lines): - line 52: // TODO: Check all videoFrames come from either CPU or GPU. - line 385: // TODO: Validate Bounds onnxruntime/core/platform/path_lib.h (2 lines): - line 177: // TODO: test if it is reg - line 208: //TODO: rewrite it with PathFindNextComponentW js/react_native/lib/backend.ts (2 lines): - line 70: // TODO: implement profiling - line 73: // TODO: implement profiling orttraining/orttraining/python/training/ortmodule/_training_manager.py (2 lines): - line 47: # TODO: Try to reuse the output buffers as some of the output tensors are same sizes, - line 55: # TODO: Non-contiguous tensor input in execution_session_run_forward, need tensor copy. onnxruntime/core/providers/cpu/tensor/unique.cc (2 lines): - line 129: // TODO: could copy blocks instead of individual items for non std::string types - line 144: // TODO: Simple copy for now. std::string would be better as std::reference_wrapper orttraining/orttraining/core/graph/gradient_builder.cc (2 lines): - line 70: // TODO: handle invalid conversion cases - line 1013: // TODO: this IA("") does not cause kernel to know it is unneeded. js/web/lib/onnxjs/instrument.ts (2 lines): - line 190: // TODO: calculate source location from 'stack' - line 242: // TODO: we want to support wildcard or regex? onnxruntime/core/providers/cpu/tensor/onehot.cc (2 lines): - line 206: // TODO potential optimization opportunity - line 207: // TODO figure out the eigen threadpool stuff for use here orttraining/orttraining/training_ops/cuda/nn/conv_grad.cc (2 lines): - line 72: // TODO: This is a cuDNN bug that gave wrong results in certain strided convolution gradient setups - line 122: // TODO: Currently we use global AlgoPerfCache for ConvGrad only. Conv's perf cache is till per node. onnxruntime/core/providers/rocm/nn/conv_transpose.cc (2 lines): - line 50: // TODO: the error message should tell which operator raises it. - line 70: // TODO: add a global cache if need to handle cases for multiple frames running simultaneously with different batch_size orttraining/orttraining/core/graph/loss_function_builder.cc (2 lines): - line 6: // TODO: solve the op version issue in the entire training framework - line 18: //TODO: Better handle op version and domain. onnxruntime/core/providers/dnnl/dnnl_node_capability.cc (2 lines): - line 332: //TODO: if initializer zero point is a vector of a unique value, we can still compute it - line 488: // TODO convert {1} outputs back to scaler {} once that is done DnnlElementwiseCapability onnxruntime/core/providers/nuphar/compiler/nuphar_compiler.cc (2 lines): - line 150: // TODO: refactor the following logic for both JIT-caching and AOT support - line 221: // TODO: change name onnxruntime/python/tools/transformers/fusion_embedlayer.py (2 lines): - line 258: # TODO: Support roberta (position starts from 2 instead of 0) in EmbedLayerNormalization kernel - line 311: # TODO: use other information (like initializer names) to identify different embedding weights automatically. orttraining/orttraining/training_ops/cuda/collective/nccl_kernels.cc (2 lines): - line 107: // TODO: temporary hack until View is improved (it doesn't work with Alias) - line 192: // TODO: temporary hack until View is improved (it doesn't work with Alias) winml/adapter/winml_adapter_c_api.h (2 lines): - line 26: // TODO: Must match onnxruntime::profiling::EventRecord - line 295: // Dml methods (TODO need to figure out how these need to move to session somehow...) js/web/lib/onnxjs/graph.ts (2 lines): - line 57: // TODO: add generic functions to manipulate the graph - line 742: // TODO: add other activation methods onnxruntime/core/providers/cuda/nn/conv.cc (2 lines): - line 55: // TODO: get maximum available size from memory areana - line 194: // TODO: Explore padding the provided input shape [N, C, D] to [N, C, 1, D] onnxruntime/core/session/provider_stubs.cc (2 lines): - line 18: // TODO: Not sure if we need to purely limit to iOS builds, so limit to __APPLE__ for now - line 85: TODO: When the NNAPI or CoreML EPs are setup to use the provider bridge the source code for that will be included orttraining/orttraining/training_ops/cuda/cuda_training_kernels.cc (2 lines): - line 93: // TODO: decprecate GatherND-1 after updating training models to opset-12 - line 278: // TODO: decprecate GatherND-1 after updating training models to opset-12 onnxruntime/core/providers/cpu/generator/range.cc (2 lines): - line 9: //TODO: fix the warnings - line 33: // TODO: Remove this contrib kernel registration and the schema from the appropriate places onnxruntime/core/optimizer/selectors_actions/selector_action_transformer.cc (2 lines): - line 90: // TODO: for now this just needs to support ONNX ops. If we ever had a transformer that was going to - line 167: // TODO: use GraphTransformer::GetCompatibleExecutionProviders if we need something more flexible onnxruntime/core/graph/graph_utils.cc (2 lines): - line 313: // TODO: Currently we remove the node and use the input name from the node being removed. - line 346: //TODO: enable the check back js/web/lib/onnxjs/backends/webgl/ops/reshape-packed.ts (2 lines): - line 93: // TODO: squeeze other shapes to 2D case - line 109: // TODO: optimize the algorithm -- in some cases, if the last two dims are js/web/lib/onnxjs/backends/webgl/ops/resize-packed.ts (2 lines): - line 80: // TODO: support other modes - line 111: // TODO:supporting other coordinateTransformModes onnxruntime/core/providers/coreml/builders/impl/resize_op_builder.cc (2 lines): - line 239: // TODO support ResizeBilinear - line 270: // TODO support ResizeBilinear orttraining/orttraining/training_ops/cuda/optimizer/adam.cc (2 lines): - line 13: // TODO: Once Schema is checked in to onnx lets fix this to match that - line 86: // TODO: temporary hack until View is improved (it doesn't work with Alias) onnxruntime/core/providers/rocm/nn/conv.cc (2 lines): - line 53: // TODO: get maximum available size from memory areana - line 193: // TODO: Remove asym padding correction. onnxruntime/python/onnxruntime_pybind_mlvalue.cc (2 lines): - line 152: // TODO: Expose knobs so that users can set fields associated with OrtArenaCfg so that we can pass it to the following method - line 205: // TODO: Expose knobs so that users can set fields associated with OrtArenaCfg so that we can pass it to the following method include/onnxruntime/core/framework/buffer_deleter.h (2 lines): - line 10: // TODO: Do we need this class or is IAllocator::MakeUniquePtr sufficient/better - line 23: // TODO: we may need consider the lifetime of alloc carefully onnxruntime/python/tools/transformers/quantize_helper.py (2 lines): - line 51: TODO: mix of in-place and return, but results are different - line 78: #TODO: inlcude external data in total model size. onnxruntime/core/providers/cpu/ml/ml_common.h (2 lines): - line 262: // TODO: Replace this with usage of code in Softmax operator - line 393: // TODO: Update TreeEnsemble* ops to use this instead of write_scores if possible. onnxruntime/core/providers/nuphar/compiler/nuphar_codegen_ctx.h (2 lines): - line 109: // TODO remove this after decoupling compiler and runtime of WeightLayout - line 112: // TODO remove this after decoupling compiler and runtime of WeightLayout onnxruntime/core/providers/rknpu/onnx_converter.cc (2 lines): - line 220: // TODO: support other type - line 399: // TODO: support it onnxruntime/wasm/api.cc (2 lines): - line 37: // TODO: This macro can be removed when we changed all APIs to return a status code. - line 95: // TODO: support profling include/onnxruntime/core/platform/ort_mutex.h (2 lines): - line 85: // TODO: is it possible to use nsync_from_time_point_ ? - line 170: // TODO: is it possible to use nsync_from_time_point_ ? onnxruntime/core/providers/nuphar/compiler/codegen_manager.h (2 lines): - line 30: // TODO add a list of condition to handle dynamic registration - line 33: // TODO: add target as an input onnxruntime/core/providers/cuda/nn/conv_transpose.cc (2 lines): - line 49: // TODO: the error message should tell which operator raises it. - line 69: // TODO: add a global cache if need to handle cases for multiple frames running simultaneously with different batch_size onnxruntime/core/providers/cpu/tensor/gather_elements.cc (2 lines): - line 148: for (int64_t i = 0; i < num_elements; ++i) { // TODO: parallelize this? didn't give any benefit in my tests - line 164: if (inner_dim_size < kParallelizationThreshold) { // TODO: tune this, arbitrary threshold onnxruntime/core/providers/nuphar/kernel.cc (2 lines): - line 46: // TODO: after AOT support, we should move it to a proper location - line 51: // TODO: rename tvm_target to a proper name onnxruntime/core/providers/nnapi/nnapi_builtin/builders/op_support_checker.cc (2 lines): - line 943: // TODO add support of scalar C for Gemm - line 1688: // TODO support 2+ inputs for Min/Max op orttraining/orttraining/python/training/ortmodule/_gradient_accumulation_manager.py (2 lines): - line 14: # TODO: enable switching the feature on/off in the middle of the training - line 71: # TODO: Make detection of parameter update robust onnxruntime/contrib_ops/cpu/attnlstm/uni_dir_attn_lstm.cc (2 lines): - line 17: //TODO: fix the warnings - line 491: // TODO: Temperately removed path: parallelize by partitioning the batch rows, onnxruntime/core/providers/cpu/ml/svmclassifier.cc (2 lines): - line 6: //TODO: fix the warnings - line 377: // TODO: Refine this rough metric to choose when to parallelize. onnxruntime/core/providers/cpu/tensor/size.cc (2 lines): - line 24: // TODO: we should not have to add the TypeConstraint below, since it is meant to be in - line 27: // TODO: Both onnxruntime and ONNX lists of types seem somewhat incomplete and incomparable. onnxruntime/core/framework/sequential_executor.cc (2 lines): - line 240: // TODO: log kernel inputs? - line 242: // TODO: log kernel outputs? orttraining/orttraining/training_ops/rocm/rocm_training_kernels.cc (2 lines): - line 89: // TODO: decprecate GatherND-1 after updating training models to opset-12 - line 241: // TODO: decprecate GatherND-1 after updating training models to opset-12 onnxruntime/core/framework/execution_frame.cc (2 lines): - line 526: // TODO: Should we re-use the block if the size is large enough? Would probably need to allow it - line 598: // TODO: differentiate reuse and alias, by add AllocKind::kAlias? onnxruntime/core/providers/dml/DmlExecutionProvider/src/AbiCustomRegistry.cpp (2 lines): - line 262: // TODO - Split apart multiple op-sets with a common domain into multiple registries, as required by Lotus - line 425: // TODO - handle non-tensor types onnxruntime/core/providers/cuda/controlflow/scan.cc (2 lines): - line 11: // TODO: It's ugly to include a .cc file but this .cc file defines the implementation of some templates which we need. - line 38: // TODO: We construct a Transpose kernel on each call as doing so is fairly lightweight. onnxruntime/core/optimizer/optimizer_execution_frame.cc (2 lines): - line 59: // TODO: node->ImplicitInputDefs() need to be added here for control flow nodes. - line 94: // TODO: node->ImplicitInputDefs() need to be added here for control flow nodes. onnxruntime/core/providers/nuphar/common/analysis/use_count_analysis.h (2 lines): - line 14: // TODO change namespace from codegen to nuphar - line 45: // TODO analysis move to namespace nuphar onnxruntime/core/providers/nuphar/mti_x86/quantize/imatmul16_extern.cc (2 lines): - line 92: // TODO: instead of calling Promote, we may consider to expose - line 126: // TODO: instead of calling Promote, we may consider to expose orttraining/orttraining/training_ops/rocm/nn/conv_grad.cc (2 lines): - line 4: //TODO Add exhaustive and default cases for algo. - line 118: // TODO: Currently we use global AlgoPerfCache for ConvGrad only. Conv's perf cache is still per node. onnxruntime/core/providers/cpu/reduction/reduction_ops.h (2 lines): - line 269: // TODO: use MLAS or BLAS - line 280: // TODO: use MLAS or BLAS orttraining/orttraining/models/runner/training_runner.cc (2 lines): - line 481: // TODO: create a list of must-to-fetch tensors and pass it to all graph transformer. - line 642: // TODO: move this to an operator in graph. onnxruntime/core/session/environment.cc (2 lines): - line 60: // TODO: Should we remove the OrtAllocatorType field from the OrtMemoryInfo struct to - line 110: // TODO should we allow sharing of non-CPU allocators? winml/lib/Api.Image/TensorToVideoFrameConverter.cpp (2 lines): - line 566: // TODO: Make an allocator for readback heaps - line 607: // TODO: Make an allocator for readback heaps js/web/lib/onnxjs/backends/webgl/ops/pool.ts (2 lines): - line 40: // TODO: support attribute 'ceil_mode' - line 118: // TODO: support attribute 'ceil_mode' and 'storage_order' onnxruntime/contrib_ops/cpu/transformers/beam_search_scorer.cc (2 lines): - line 14: // Could reduce the chance of arithmetic overflow. TODO: fix it - line 32: // TODO: when T is FP16, compute in FP32, then cast result back to FP16. length_penalty_ might also be float. orttraining/orttraining/python/training/ortmodule/torch_cpp_extensions/__init__.py (2 lines): - line 25: TODO: Create a generic mechanism to pass arguments from ORTModule into each extension setup.py - line 26: TODO: Create environment variables to allow extensions to be hosted outside ONNX runtime installation folder onnxruntime/core/framework/orttraining_partial_executor.cc (2 lines): - line 239: // TODO: log kernel inputs? - line 257: // TODO: log kernel outputs? onnxruntime/core/providers/cuda/shared_inc/fpgeneric.h (2 lines): - line 11: // NV_TODO: optimize speed -- pass things needed in, optimize kernel speed, add half2 - line 12: // NV_TODO: investigate cub support for half onnxruntime/core/framework/allocatormgr.h (2 lines): - line 16: // TODO: update OrtMemoryInfo, use unordered_set instead - line 43: // TODO: Only used for TRT and CUDA EP currently, need to add more identifiers to use it across all EPs onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc (2 lines): - line 628: // TODO: this is actually used for the cuda kernels which explicitly ask for inputs from CPU. - line 1416: // TODO: remove default capture onnxruntime/core/framework/iexecutor.h (1 line): - line 40: // TODO: as fetch_allocators is optional, it should be a pointer instead of reference orttraining/orttraining/core/graph/loss_function_registry.cc (1 line): - line 36: // TODO: support setting attributes of the custom op. onnxruntime/core/providers/cpu/math/einsum_utils/einsum_auxiliary_ops.cc (1 line): - line 79: // TODO: Parallelize this operation onnxruntime/python/tools/transformers/onnx_model_bert_keras.py (1 line): - line 76: # TODO: shall we add back the checking of children op types. orttraining/orttraining/python/training/ortmodule/torch_cpp_extensions/cpu/aten_op_executor/aten_op_executor.cc (1 line): - line 99: default: // TODO: will add more type support if needed. onnxruntime/python/tools/transformers/profiler.py (1 line): - line 332: # TODO: shall we have a separated group for nodes with subgraph? tools/python/check_model_can_use_ort_mobile_pkg.py (1 line): - line 162: # TODO: Would be nice to specify ORT release and pull the config for that release. js/common/lib/env-impl.ts (1 line): - line 14: // TODO standadize the getter and setter convention in env for other fields. onnxruntime/core/framework/tensor_type_and_shape.cc (1 line): - line 328: // TODO: This is consistent with the previous implementation but inconsistent with GetValueType which returns onnxruntime/core/providers/cpu/ml/linearregressor.cc (1 line): - line 98: // TODO: Add support for 'double' to the scoring functions in ml_common.h orttraining/orttraining/core/framework/pipeline.cc (1 line): - line 14: //TODO: fix the warnings onnxruntime/core/providers/cuda/object_detection/non_max_suppression_impl.cu (1 line): - line 27: //TODO:fix the warnings onnxruntime/core/common/profiler.h (1 line): - line 138: * TODO: Consider MEMFS or IndexedDB instead of console. onnxruntime/core/providers/dnnl/subgraph/dnnl_subgraph_primitive.cc (1 line): - line 518: //TODO: expand to arbitrary permutation or transpose on given 2 dims for higher dimensional tensors cmake/onnxruntime_mlas.cmake (1 line): - line 34: #TODO: set MASM flags properly onnxruntime/core/providers/nuphar/compiler/x86/op_ir_creator/nn/pool_ops.cc (1 line): - line 40: // TODO: handle symbolic dimensions onnxruntime/core/providers/acl/nn/conv.cc (1 line): - line 293: //TODO: valildate shapes onnxruntime/core/providers/nuphar/runtime/compute_ctx.h (1 line): - line 23: // TODO change name space from tvm_codegen to nuphar onnxruntime/core/framework/random_seed.cc (1 line): - line 14: //TODO: Fix the warning. The variable should be put in the environment class. onnxruntime/core/providers/cpu/rnn/lstm_base.cc (1 line): - line 6: //TODO: fix the warnings server/grpc/prediction_service_impl.cc (1 line): - line 16: //TODO: (csteegz) Add modelspec for both paths. onnxruntime/contrib_ops/cpu/transformers/beam_search_scorer.h (1 line): - line 121: // TODO: use ORT allocator to avoid allocating from heap directly js/web/lib/onnxjs/backends/webgl/ops/fuse-utils.ts (1 line): - line 28: // TODO: adding other activations that can be fused. onnxruntime/contrib_ops/cpu/transformers/sequences.cc (1 line): - line 3: // Could reduce the chance of arithmetic overflow. TODO: fix it orttraining/orttraining/training_ops/rocm/math/div_grad.cc (1 line): - line 59: // TODO: BinaryElementwiseBroadcastPrepare shall take dy_tensor as const Tensor*. onnxruntime/python/tools/quantization/onnx_quantizer.py (1 line): - line 202: # TODO: convert it to the specified input_type onnxruntime/core/graph/dml_ops/dml_defs.cc (1 line): - line 131: // TODO in training mode, it may be possible to infer some of tools/ci_build/github/azure-pipelines/templates/win-gpu-ci.yml (1 line): - line 223: #TODO: move this step to a CPU-only machine to save GPU resources. orttraining/tools/scripts/pipeline_model_split.py (1 line): - line 376: # TODO: move this info to a file (json?) and load the data from there. onnxruntime/core/providers/cuda/tensor/nonzero_impl.cu (1 line): - line 14: //TODO:check overflow orttraining/orttraining/core/graph/gradient_builder.h (1 line): - line 8: // TODO: maybe group the gradient builders and split them into different files. winml/lib/Api.Ort/OnnxruntimeEngine.cpp (1 line): - line 712: // TODO: what is the difference between the device allocator and the arena allocator? onnxruntime/core/providers/cpu/nn/conv_attributes.h (1 line): - line 59: // TODO: Re-enable when attributes values are guaranteed to be filled. onnxruntime/core/providers/cuda/reduction/reduction_ops.cc (1 line): - line 154: // TODO ReduceKernel::ReduceKernelShared() is still used by some other training classes though it's not used here - this should be refactored. onnxruntime/core/providers/cuda/reduction/reduction_ops.h (1 line): - line 51: // TODO: Clean up the CUDAExecutionProvider interface to avoid this onnxruntime/core/providers/rocm/reduction/reduction_ops.cc (1 line): - line 151: // TODO ReduceKernel::ReduceKernelShared() is still used by some other training classes though it's not used here - this should be refactored. js/web/lib/onnxjs/backends/webgl/inference-handler.ts (1 line): - line 138: // TODO: currently this is a hacking to overwrite Conv's weight. The correct way to do this should be: onnxruntime/core/optimizer/gemm_sum_fusion.h (1 line): - line 20: TODO: Support the Add use case: Sum(x, y) ~= Add. onnxruntime/core/providers/cpu/optional/optional_ops.cc (1 line): - line 106: // TODO: Should we also ensure that element type for tensor is set ? orttraining/orttraining/training_ops/cuda/math/div_grad.h (1 line): - line 12: class DivGrad : public ReduceKernel { // TODO: not to derive from ReduceKernel. onnxruntime/core/providers/nuphar/nuphar_provider_factory.cc (1 line): - line 8: //#include "core/codegen/passes/utils/codegen_context.h" // TODO: remove it onnxruntime/core/providers/cpu/tensor/pad.cc (1 line): - line 24: // TODO: Remove this contrib kernel registration and the schema from the appropriate places include/onnxruntime/core/framework/allocator.h (1 line): - line 64: // TODO: Find a better name than Reserve() and update in all places. onnxruntime/core/providers/nuphar/runtime/sequential/basic.cc (1 line): - line 57: // TODO: change it to use provider::CopyTensor for non-CPU devices onnxruntime/core/providers/rocm/rocm_execution_provider.cc (1 line): - line 2206: // TODO: this is actually used for the rocm kernels which explicitly ask for inputs from CPU. onnxruntime/core/providers/nnapi/nnapi_builtin/builders/model_builder.cc (1 line): - line 225: // TODO: support other type onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.cc (1 line): - line 80: // TODO: Task 812756: NNAPI EP, add support for subgraph (If and Loop operators) onnxruntime/core/providers/cuda/math/einsum.h (1 line): - line 19: // TODO: Clean up the CUDAExecutionProvider interface to avoid this onnxruntime/core/platform/posix/env.cc (1 line): - line 237: // TODO if you need the number of physical cores you'll need to parse onnxruntime/core/providers/nuphar/compiler/traverse_shape_infer.cc (1 line): - line 10: // TODO retire this file onnxruntime/contrib_ops/cpu/maxpool_with_mask.h (1 line): - line 203: // TODO: fix this checker later onnxruntime/core/providers/cuda/math/einsum_utils/einsum_auxiliary_ops.cc (1 line): - line 27: // TODO: Currently, triggers copy on stream 0, investigate if we can still do that orttraining/orttraining/core/framework/gradient_graph_builder.cc (1 line): - line 285: //TODO: might not need two sets, the union of them might be enough onnxruntime/core/providers/cuda/cuda_allocator.h (1 line): - line 53: //TODO: add a default constructor onnxruntime/core/providers/nnapi/nnapi_builtin/model.h (1 line): - line 91: // TODO: include/onnxruntime/core/framework/func_api.h (1 line): - line 3: using onnxruntime::common::Status; // TODO: Needed by WinML, but shouldn't be put into the global namespace like this onnxruntime/core/framework/kernel_registry.cc (1 line): - line 173: // TODO: We should check that names specified in kernel_type_constraints are onnxruntime/core/mlas/lib/convolve.cpp (1 line): - line 1227: // TODO: support more general depthwise convolution. onnxruntime/core/framework/mem_pattern_planner.h (1 line): - line 36: // TODO: OverlappingTimeSchedules should be private onnxruntime/python/tools/transformers/convert_beam_search.py (1 line): - line 200: # TODO: Use external format if input has extra data. onnxruntime/core/optimizer/gemm_sum_fusion.cc (1 line): - line 138: // TODO: verify if scalar shape works here together with matmul_add_fusion. onnxruntime/core/providers/cpu/math/det.cc (1 line): - line 6: //TODO: fix the warnings tools/python/util/ort_format_model/types.py (1 line): - line 54: # TODO: Decide if we need to wrap the type in a std::vector. Issue is that the element type is internal onnxruntime/core/framework/execution_frame.h (1 line): - line 139: // TODO: These two AllocateMLValue... methods are in the API purely for unit test usage. onnxruntime/core/providers/cuda/tensor/compress_impl.cu (1 line): - line 9: //TODO:fix the warnings onnxruntime/core/providers/cpu/math/matmul.cc (1 line): - line 113: // TODO: replace it with GemmBatch for performance, it's OK for now as GemmBatch unrolls as well onnxruntime/python/tools/quantization/onnx_model.py (1 line): - line 308: # TODO:use OnnxModel.graph_topological_sort(self.model.graph) from transformers.onnx_model onnxruntime/core/providers/nuphar/compiler/x86/scheduler/tensorize/tensorize_base.h (1 line): - line 7: // TODO move this tvm_codegen onnxruntime/python/onnxruntime_inference_collection.py (1 line): - line 322: self._model_bytes = path_or_bytes # TODO: This is bad as we're holding the memory indefinitely onnxruntime/core/providers/cpu/generator/random.cc (1 line): - line 255: std::uniform_real_distribution dist(0.0, 1.0); // TODO: should this be initialized per batch? onnxruntime/core/providers/nuphar/compiler/nuphar_op_ir_builder.cc (1 line): - line 38: // TODO: add support for sequence not axis 0 onnxruntime/core/providers/acl/math/gemm.h (1 line): - line 138: //TODO: valildate shapes onnxruntime/core/providers/nuphar/common/analysis/subgraph_partition_stats.cc (1 line): - line 11: // TODO: Add memory analysis orttraining/orttraining/training_ops/cuda/communication/send.cc (1 line): - line 160: // TODO move the following variables to member variables for extending life-time onnxruntime/core/framework/prepacked_weights_container.cc (1 line): - line 18: // TODO: Investigate benefits of using an arena based allocator onnxruntime/core/framework/compute_capability.h (1 line): - line 15: // TODO: if there is a FusedKernelFn attached, onnxruntime will generate onnxruntime/core/optimizer/fast_gelu_fusion.cc (1 line): - line 292: // TODO: re-use node arg of mul5 so that it is allowed to be graph output (Need modify CheckNode as well). orttraining/orttraining/python/training/ortmodule/_io.py (1 line): - line 51: TODO: Remove once PyTorch 1.8.2 or newer is released onnxruntime/core/optimizer/selectors_actions/actions.cc (1 line): - line 39: // TODO: It's slightly insane we don't support optionally removing the output edges as part of Graph::RemoveNode winml/lib/Api/ImageFeatureDescriptor.cpp (1 line): - line 128: // TODO: Need to add denotations here onnxruntime/core/providers/dnnl/subgraph/dnnl_matmul.cc (1 line): - line 73: //TODO: use format any to choose the best layout onnxruntime/core/providers/nuphar/compiler/x86/scheduler/ort_type_schedule.cc (1 line): - line 152: // TODO: remove compute_root onnxruntime/contrib_ops/cpu/crop_and_resize.cc (1 line): - line 25: //TODO: fix the warnings onnxruntime/core/providers/nuphar/compiler/x86/op_ir_creator/math/matmul.cc (1 line): - line 18: // TODO: remove tvm core function onnxruntime/core/providers/rocm/cu_inc/common.cuh (1 line): - line 314: // TODO ROCM added support recently, should verify. orttraining/orttraining/eager/opgen/opgen/atenops.py (1 line): - line 97: # TODO: this is a temporary allowlist for ops need type promotion onnxruntime/core/common/profiler.cc (1 line): - line 87: //TODO: sync_gpu if needed. onnxruntime/core/providers/cuda/cuda_call.cc (1 line): - line 96: //TODO: avoid using const_cast onnxruntime/core/providers/coreml/builders/impl/builder_utils.cc (1 line): - line 102: // TODO: support other type onnxruntime/core/providers/coreml/coreml_execution_provider.cc (1 line): - line 50: // TODO investigate whether we want to support subgraph using CoreML EP tools/python/gen_contrib_doc.py (1 line): - line 142: # TODO: link to the Changelog.md onnxruntime/core/providers/cuda/math/softmax.h (1 line): - line 49: // TODO: Clean up the CUDAExecutionProvider interface to avoid this onnxruntime/core/providers/cpu/ml/scaler.cc (1 line): - line 90: if (x_size < kParallelizationThreshold) { // TODO: tune this, arbitrary threshold winml/adapter/winml_adapter_session.cpp (1 line): - line 58: // TODO Instead of returning an error, should we set mem pattern to false here and log a warning saying so? orttraining/orttraining/training_ops/cuda/math/div_grad.cc (1 line): - line 59: // TODO: BinaryElementwiseBroadcastPrepare shall take dy_tensor as const Tensor*. csharp/src/Microsoft.ML.OnnxRuntime/Tensors/DenseTensor.shared.cs (1 line): - line 165: // TODO: use Span.IndexOf when/if it removes the IEquatable type constraint orttraining/orttraining/core/optimizer/graph_transformer_utils.cc (1 line): - line 193: // TODO hack - constant folding currently doesn't work after mixed precision transformation so it's disabled for now onnxruntime/core/platform/windows/logging/etw_sink.cc (1 line): - line 105: // TODO: Validate if this filtering makes sense. onnxruntime/core/codegen/mti/tensor/cast_ops.cc (1 line): - line 26: // TODO: remove it, after TVM is fixed onnxruntime/contrib_ops/cuda/bert/longformer_attention.cc (1 line): - line 68: // TODO: only calculate once per model. onnxruntime/core/providers/dml/DmlExecutionProvider/src/MLOperatorAuthorImpl.h (1 line): - line 641: // TODO - ... onnxruntime/contrib_ops/cuda/bert/attention.cc (1 line): - line 81: // TODO: use custom kernel of expand to improve the performance. onnxruntime/core/providers/cpu/object_detection/roialign.cc (1 line): - line 52: //TODO: fix the warnings orttraining/orttraining/training_ops/cpu/communication/send.cc (1 line): - line 80: // TODO move the following variables to member variables for extending life-time onnxruntime/core/providers/cpu/rnn/rnn.cc (1 line): - line 12: //TODO: fix the warnings orttraining/orttraining/models/runner/training_runner.h (1 line): - line 146: // TODO support more algos when they become available. orttraining/orttraining/eager/ort_tensor.cpp (1 line): - line 81: // TODO: wrap with change generation guard onnxruntime/core/framework/parallel_executor.cc (1 line): - line 124: // TODO: Convert RunNodeAsync return Status. onnxruntime/core/optimizer/transpose_optimizer/transpose_optimizer.cc (1 line): - line 1531: // TODO: check binary size of this and replace it with constexpr if large onnxruntime/python/tools/transformers/onnx_model_bert.py (1 line): - line 266: # TODO: merge ConstantOfShape -->Cast to ConstantOfShape (need update the data type of value) onnxruntime/core/providers/nuphar/compiler/nuphar_schedule_builder.cc (1 line): - line 14: // TODO change name space orttraining/orttraining/training_ops/cuda/communication/nccl_service.cc (1 line): - line 230: // TODO: Create a general class to describe for computation topology and unify all similar uses. onnxruntime/core/optimizer/attention_fusion_helper.h (1 line): - line 1332: TODO: replace Gemm_Subgraph by MatMul + Add onnxruntime/core/providers/nnapi/nnapi_builtin/builders/model_builder.h (1 line): - line 32: /* TODO support these options server/http/core/session.cc (1 line): - line 20: // TODO: make the max request size configable. onnxruntime/core/dll/dllmain.cc (1 line): - line 27: //TODO: Don't do it when Protobuf_USE_STATIC_LIBS is OFF onnxruntime/core/optimizer/selectors_actions/actions.h (1 line): - line 9: #include "core/graph/graph_utils.h" // TODO: Minimize usage of this given we want to use Actions in a minimal build tools/python/ort_test_dir_utils.py (1 line): - line 55: # TODO: see if we need to provide a way to specify these values. could ask for the whole orttraining/orttraining/training_ops/cuda/tensor/gather_nd_grad.cc (1 line): - line 72: // TODO this memset can be expensive, a sparse tensor representation would help here onnxruntime/core/session/inference_session_utils.cc (1 line): - line 169: // TODO: Support all valid session options onnxruntime/core/providers/cuda/tensor/gather_nd.cc (1 line): - line 123: // TODO: decprecate GatherND-1 after updating training models to opset-12 orttraining/orttraining/training_ops/cpu/nn/dropout_7.cc (1 line): - line 45: // TODO: Compute is a const function, generator cannot be a private meber orttraining/orttraining/python/training/ortmodule/torch_cpp_extensions/cuda/torch_gpu_allocator/setup.py (1 line): - line 14: # TODO: Implement a cleaner way to auto-generate torch_gpu_allocator.cc winml/adapter/winml_adapter_c_api.cpp (1 line): - line 63: // Dml methods (TODO need to figure out how these need to move to session somehow...) onnxruntime/python/tools/transformers/benchmark_gpt2.py (1 line): - line 211: use_external_data_format = (config.n_layer > 24) #TODO: find a way to check model size > 2GB onnxruntime/python/tools/pytorch_export_contrib_ops.py (1 line): - line 84: # TODO: replace this once PyTorch supports unregister natively. onnxruntime/core/framework/ort_value_tensor_slicer.cc (1 line): - line 81: // TODO: Ideally we could avoid the overhead of creating a new Tensor (mainly cost of copying type and shape info) onnxruntime/contrib_ops/cpu/quantization/qembed_layer_norm.cc (1 line): - line 93: // TODO: Profile and tune this batch parallel execution based on input size. orttraining/orttraining/core/framework/tensorboard/event_writer.cc (1 line): - line 33: // TODO: get hostname for inclusion in Tensorboard events filename tools/ci_build/upload_python_package_to_azure_storage.py (1 line): - line 74: # TODO: figure out a way to secure args.account_key to prevent later code changes onnxruntime/contrib_ops/cpu/attnlstm/attention_wrapper.cc (1 line): - line 11: //TODO: fix the warnings onnxruntime/core/framework/memory_info.h (1 line): - line 220: //TODO: The dynamic and statically planned alignments may not be the same, need to check js/web/lib/onnxjs/backends/webgl/ops/flatten.ts (1 line): - line 35: // TODO: Support string type include/onnxruntime/core/optimizer/graph_transformer_utils.h (1 line): - line 31: TODO: This is visible for testing at the moment, but we should rather make it private. */ onnxruntime/core/providers/cpu/rnn/uni_directional_lstm.cc (1 line): - line 7: //TODO: fix the warnings onnxruntime/core/providers/cpu/rnn/rnn.h (1 line): - line 29: //TODO: is it optional or not? onnxruntime/python/onnxruntime_pybind_iobinding.cc (1 line): - line 40: // TODO: Support non-tensors orttraining/orttraining/python/training/optim/config.py (1 line): - line 63: # TODO: monitor this for perf issues onnxruntime/contrib_ops/cuda/math/bias_softmax_impl.cu (1 line): - line 276: //TODO: fast_divmod only supports int32 onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc (1 line): - line 17: //TODO: fix the warnings orttraining/orttraining/training_ops/cpu/nn/dropout_7.h (1 line): - line 17: // TODO: enable following when is_train is present onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.h (1 line): - line 20: // TODO: add more parameter in consructor to support different target csharp/src/Microsoft.ML.OnnxRuntime/NativeMethods.shared.cs (1 line): - line 256: // TODO: Make this save the pointer, and not copy the whole structure across onnxruntime/python/tools/transformers/gpt2_parity.py (1 line): - line 193: ) #TODO: shall we use one-sided: less or greater according to "top1_match_rate" onnxruntime/python/tools/transformers/benchmark_helper.py (1 line): - line 36: # TODO: Add more. onnxruntime/core/providers/nuphar/kernel.h (1 line): - line 49: // TODO: remove it after weight layout refactoring onnxruntime/core/providers/cpu/nn/batch_norm_helper.h (1 line): - line 11: //TODO: fix the warnings onnxruntime/core/providers/cpu/math/element_wise_ops.h (1 line): - line 14: // TODO: fix the warnings cmake/onnxruntime_graph.cmake (1 line): - line 84: #TODO: the graph library should focus on ONNX IR, it shouldn't depend on math libraries like MKLML/OpenBlas onnxruntime/core/graph/graph_viewer.cc (1 line): - line 249: // TODO: See if we need to calculate the root_nodes_ of the filtered graph. orttraining/orttraining/python/training/ortmodule/_custom_autograd_function_exporter.py (1 line): - line 126: # TODO: add fully-qualified name. cmake/onnxruntime_objectivec.cmake (1 line): - line 89: CXX_STANDARD 17 # TODO remove when everything else moves to 17 onnxruntime/core/common/path_utils.h (1 line): - line 14: * TODO: add support for arguments of type std::wstring. Currently it is not supported as the underneath orttraining/orttraining/training_ops/cuda/communication/recv.cc (1 line): - line 166: // TODO move the following variables to member variables for extending life-time onnxruntime/core/optimizer/slice_elimination.cc (1 line): - line 118: // TODO: Take into account the input's shape to get a tighter bound for the ends. js/web/lib/onnxjs/backends/webgl/ops/depth-to-space.ts (1 line): - line 73: // TODO: Support string depth-to-space. js/common/lib/tensor-utils.ts (1 line): - line 22: // TODO: add more tensor utilities onnxruntime/contrib_ops/cpu/transformers/beam_search_parameters.cc (1 line): - line 43: // TODO: limit num_beams > 1 when we can have another operator for greedy search. onnxruntime/core/providers/cuda/math/topk_impl.cu (1 line): - line 12: //TODO:fix the warnings onnxruntime/core/providers/cpu/tensor/gather_nd.cc (1 line): - line 13: // TODO: Remove this contrib kernel registration and the schema from the appropriate places onnxruntime/core/graph/function_container.h (1 line): - line 8: //TODO: we need to make it a stand-alone header because both graph.cc and model.cc need to implement create instance of the graph object. onnxruntime/core/providers/nuphar/compiler/codegen_manager.cc (1 line): - line 243: // TODO Add isa support onnxruntime/core/providers/coreml/builders/model_builder.cc (1 line): - line 167: // TODO: support other type onnxruntime/core/codegen/common/creator.h (1 line): - line 7: // TODO rename this file to creator_base include/onnxruntime/core/graph/basic_types.h (1 line): - line 37: // TODO - Evaluate switching the types below to support transparent comparators and enable onnxruntime/core/providers/cuda/tensor/scatter_nd.cc (1 line): - line 50: // TODO: Run benchmarks to determine if a dedicated kernel doing data copy will be faster than invoking cudaMemcpy ? onnxruntime/python/tools/tensorrt/perf/benchmark.py (1 line): - line 82: #command.extend([inputs_arg]) TODO: rebind IO inputs in TRT 8.2 onnxruntime/core/providers/nuphar/runtime/sequential/loop.cc (1 line): - line 8: // TODO: refactor it orttraining/orttraining/core/graph/optimizer/lamb_optimizer_builder.cc (1 line): - line 230: //TODO: need to support float -> float16 and float16-> float conversion onnxruntime/core/providers/cpu/controlflow/loop.cc (1 line): - line 610: // TODO: We could try and call ExecuteGraph to get the output shape from fetches so the rank is correct, csharp/src/Microsoft.ML.OnnxRuntime/DisposableNamedOnnxValue.shared.cs (1 line): - line 131: /* Get Tensor element type */ //TODO: Assumed value is Tensor, need to support non-tensor types in future winml/adapter/winml_adapter_apis.h (1 line): - line 62: // Dml methods (TODO need to figure out how these need to move to session somehow...) onnxruntime/core/framework/session_options.h (1 line): - line 90: unsigned max_num_graph_transformation_steps = 10; // TODO choose a good default here? onnxruntime/core/providers/nuphar/scripts/model_quantizer.py (1 line): - line 110: # TODO: support GEMM op if needed onnxruntime/__init__.py (1 line): - line 41: # TODO: thiagofc: Temporary experimental namespace for new PyTorch front-end onnxruntime/core/optimizer/selectors_actions/helpers.cc (1 line): - line 86: // TODO: currently variadic inputs have their count corrected (should be one entry with the total number of include/onnxruntime/core/framework/alloc_kind.h (1 line): - line 13: // caller. TODO: Make sure this semantics is clear in InferenceSession API. orttraining/orttraining/core/graph/optimizer/adam_optimizer_builder.cc (1 line): - line 96: //TODO: need to support float -> float16 and float16-> float conversion onnxruntime/core/providers/dml/DmlExecutionProvider/src/GraphKernelHelper.cpp (1 line): - line 324: // TODO: This is a hack which strips the suffix added within Lotus transforms that insert mem copies. java/src/main/java/ai/onnxruntime/OnnxRuntime.java (1 line): - line 296: // TODO: Switch this to Path.of when the minimum Java version is 11. onnxruntime/core/providers/rocm/rocm_call.cc (1 line): - line 99: //TODO: avoid using const_cast onnxruntime/python/tools/quantization/operators/pad.py (1 line): - line 52: # TODO: check quantize_inputs after sub graph is supported csharp/src/Microsoft.ML.OnnxRuntime/OrtValue.shared.cs (1 line): - line 279: //TODO: not sure how to support reverse stride. may be able to calculate the shape differently include/onnxruntime/core/common/logging/logging.h (1 line): - line 75: // TODO: What other high level categories are meaningful? Model? Optimizer? Execution? onnxruntime/core/framework/execution_provider.cc (1 line): - line 96: // TODO: Should we remove the OrtAllocatorType field from OrtMemoryInfo to orttraining/orttraining/core/framework/gradient_graph_builder.h (1 line): - line 44: {"Pow", {1}}, // TODO: Pow's input_1 is differentiable, but gradient not yet implemented onnxruntime/core/providers/cpu/reduction/reduction_ops.cc (1 line): - line 7: //TODO: fix the warnings onnxruntime/core/providers/rocm/reduction/reduction_ops.h (1 line): - line 51: // TODO: Clean up the ROCMExecutionProvider interface to avoid this onnxruntime/core/providers/cpu/nn/lrn.cc (1 line): - line 24: //TODO: fix the warnings include/onnxruntime/core/platform/EigenNonBlockingThreadPool.h (1 line): - line 593: // TODO: Revisit whether there is a significant benefit for the current tools/python/FindOptimizerOpsetVersionUpdatesRequired.py (1 line): - line 48: # TODO: handle automatically by merging lines cmake/onnxruntime_python.cmake (1 line): - line 72: #TODO: fix the warnings onnxruntime/core/providers/cpu/activation/activations.h (1 line): - line 21: // TODO: Tune the cost server/get_boost.cmake (1 line): - line 37: # TODO: let user give their own Boost installation onnxruntime/core/providers/cuda/object_detection/non_max_suppression.cc (1 line): - line 49: // TODO: use cub::DeviceSegmentedRadixSort::SortPairsDescending instead of cub::DeviceRadixSort::SortPairsDescending onnxruntime/contrib_ops/cuda/bert/skip_layer_norm.cc (1 line): - line 112: static_cast(element_count), //TODO: check range onnxruntime/core/session/inference_session.h (1 line): - line 533: * @remarks TODO: Provide way to load from in-memory bytes without copying. InferenceSession would need to orttraining/pytorch_frontend_examples/mnist_training.py (1 line): - line 54: # TODO: comple this once ORT training can do evaluation. onnxruntime/core/framework/ort_value_name_idx_map.h (1 line): - line 12: // TODO: this is a static hash lookup, it's easy to do it better onnxruntime/python/tools/quantization/operators/attention.py (1 line): - line 23: # TODO This is a temporary fix to stop exporting QAttention with qkv_hidden_sizes orttraining/orttraining/eager/ort_backends.h (1 line): - line 39: //TODO: we might want to support load custom op schema on the fly onnxruntime/core/providers/cpu/math/top_k.cc (1 line): - line 172: // TODO: May want a different calculation for each branch below instead. onnxruntime/contrib_ops/cpu/attnlstm/bahdanau_attention.cc (1 line): - line 11: // TODO: fix the warnings orttraining/orttraining/python/training/ortmodule/__init__.py (1 line): - line 73: # TODO: detect when installed extensions are outdated and need reinstallation. Hash? Version file? onnxruntime/python/tools/transformers/fusion_gpt_attention_no_past.py (1 line): - line 22: # TODO: detect num_heads from graph like FusionAttention onnxruntime/core/providers/nuphar/compiler/nuphar_schedule_builder.h (1 line): - line 10: // TODO change name space onnxruntime/python/onnxruntime_validation.py (1 line): - line 125: # TODO: rcom onnxruntime/python/tools/transformers/fusion_layernorm.py (1 line): - line 217: #TODO: add epsilon attribute onnxruntime/core/codegen/mti/nn/lstm.cc (1 line): - line 106: // TODO: handle more general cases for activations f, h, g and activation_alpha and orttraining/orttraining/training_ops/cpu/communication/recv.cc (1 line): - line 81: // TODO move the following variables to member variables for extending life-time include/onnxruntime/core/session/environment.h (1 line): - line 16: /** TODO: remove this class onnxruntime/python/tools/transformers/fusion_gpt_attention.py (1 line): - line 302: # TODO: match mask path for GPT2LMHeadModel_BeamSearchStep. onnxruntime/core/common/cpuid_info.cc (1 line): - line 114: // TODO implement hardware feature detection in windows. include/onnxruntime/core/framework/tensor.h (1 line): - line 24: //TODO:ensure dtype_!=nullptr onnxruntime/core/providers/nuphar/compiler/x86/scheduler/tensorize/tensorize_utilities.h (1 line): - line 7: // TODO move this tvm_codegen onnxruntime/contrib_ops/cpu/bert/attention_cpu_base.h (1 line): - line 12: //TODO: fix the warnings onnxruntime/core/mlas/lib/pooling.cpp (1 line): - line 1288: // TODO: change to use MlasExecuteThreaded onnxruntime/core/providers/cpu/object_detection/non_max_suppression.cc (1 line): - line 18: //TODO:fix the warnings onnxruntime/core/providers/cpu/tensor/concat.cc (1 line): - line 218: // TODO: Minor Optimization possibility: This input_axis_patch will be common across all inputs onnxruntime/core/providers/cpu/nn/shrink.cc (1 line): - line 34: //TODO: fix the warnings server/http/predict_request_handler.h (1 line): - line 15: // TODO: decide whether this should be a class server/serializing/tensorprotoutils.cc (1 line): - line 154: // TODO: complex64 complex128 onnxruntime/core/providers/nuphar/common/analysis/subgraph_partition_stats.h (1 line): - line 13: // TODO: rename class name to more target-specific in the tvm refactoring onnxruntime/core/codegen/mti/nn/lstm.h (1 line): - line 9: // TODO: refactor the LSTMcell building to a tvm function onnxruntime/core/providers/migraphx/hip_allocator.h (1 line): - line 53: //TODO: add a default constructor onnxruntime/core/framework/session_state_utils.cc (1 line): - line 239: // TODO: if the tensor need be copied, does it have enough room? onnxruntime/contrib_ops/cpu/bert/attention_helper.h (1 line): - line 105: // TODO: mask_index can be used in softmax to save some calculation. onnxruntime/python/backend/backend.py (1 line): - line 68: # is generating attribute error. TODO investigate the pipelines to onnxruntime/core/providers/cpu/rnn/rnn_helpers.cc (1 line): - line 20: //TODO: fix the warnings onnxruntime/core/framework/parallel_executor.h (1 line): - line 63: // TODO: Temporary threadpool for the executor. This is a costly way to handle the problem. js/web/lib/onnxjs/attribute.ts (1 line): - line 259: // TODO: Subgraph not supported yet. onnxruntime/python/tools/symbolic_shape_infer.py (1 line): - line 1014: assert not is_sequence(vi.type) # TODO: handle loop accumulation in sequence_type orttraining/orttraining/core/framework/communication/mpi/mpi_context.cc (1 line): - line 4: #define SHARED_PROVIDER_TODO 0 winml/lib/Api/LearningModelBinding.cpp (1 line): - line 248: // TODO: this format for unbound output needs more discussion onnxruntime/core/framework/TensorSeq.h (1 line): - line 83: // TODO: optimization opportunity - if all tensors in the seq are scalars, we can potentially represent them onnxruntime/core/optimizer/graph_transformer_mgr.h (1 line): - line 37: // TODO: remove this when we update to 6.1 or later onnxruntime/core/framework/func_kernel.h (1 line): - line 28: //TODO: we are only provide host allocate method in compute context. tools/ci_build/github/azure-pipelines/templates/mac-ci.yml (1 line): - line 70: # TODO: ESRP team is working on enable signing workflow on Mac. Should enable the following step when it's ready. onnxruntime/core/providers/cpu/tensor/space_depth_ops.cc (1 line): - line 4: // TODO: x86 release builds produce warning C4723: potential divide by 0 orttraining/orttraining/training_ops/cpu/nn/conv_grad.cc (1 line): - line 37: // TODO: validataion might not be needed, since it's already done once in the fw pass onnxruntime/core/providers/nuphar/common/nuphar_tvm_utils.cc (1 line): - line 161: // TODO: force checksum validation in final release onnxruntime/core/providers/cpu/sequence/sequence_ops.cc (1 line): - line 18: // TODO: The current implementation of sequence ops relies on tensor copies. Ideally we should try to avoid onnxruntime/core/framework/session_state.h (1 line): - line 435: // TODO: The CUDA EP may not need to use the per-thread allocator for allocations that would use this map onnxruntime/core/graph/model.cc (1 line): - line 161: // TODO: Check if we can upgrade all the current opset 6 models that are being tested onnxruntime/contrib_ops/cuda/quantization/attention_quantization.cc (1 line): - line 185: nullptr, // TODO: support add_qk in quantized attention onnxruntime/python/tools/quantization/calibrate.py (1 line): - line 450: # TODO: Currently we have different collect() for entropy and percentile method respectively. onnxruntime/contrib_ops/cpu/attnlstm/deep_cpu_attn_lstm.cc (1 line): - line 13: //TODO: fix the warnings onnxruntime/core/providers/nuphar/common/analysis/shape_expr.h (1 line): - line 7: // TODO retire this file onnxruntime/core/providers/nuphar/partition/graph_partitioner.cc (1 line): - line 151: // TODO: change this logic to removing a partition with only all alias ops onnxruntime/core/codegen/mti/nn/pool_ops.cc (1 line): - line 14: // TODO: topi only support 2d-pool, MaxPool1d and MaxPool3d will need to be added if necessary. onnxruntime/core/providers/rocm/rocm_allocator.h (1 line): - line 53: //TODO: add a default constructor orttraining/orttraining/python/training/ortmodule/_inference_manager.py (1 line): - line 37: # TODO: Try to reuse the output buffers as some of the output tensors are same sizes,