Summary: 2361 instances, 2030 unique

Text	Count
// TODO: Less copy/paste between this and normal dicts.	1
// TODO: https://github.com/pytorch/pytorch/issues/47442	1
// TODO: propagate exception_ptr to the caller side	1
# TODO: is there a way to split by device and dtype without appending in the inner loop?	1
// TODO static_assert(AllowDeprecatedTypes, "You tried to register a kernel with an unsupported output type: std::vector<T>. Please use List<T> instead.");	1
// TODO: we don't actually need separate instantiations per dtype;	1
// TODO: currently we do not extract a common factor if divisor and	1
// TODO Avoid vector allocation. One idea would be to keep the std::vector	1
// TODO: Optimize somehow. Remember iterator instead of searching for it linearly.	1
# FIXME this may have the wrong shape when support contains batched	1
# TODO: handle module -> add_scalar -> add_scalar	1
// TODO: support runtime state capture similar to `jitted_gpu_kernel`.	1
# FIXME: Unfortunately, for Windows, we are missing a worker	1
# TODO: grad_output size asserts in THNN	1
# TODO: Make sure out argument is guaranteed to be self	1
# TODO: need to extend this to consider all relevant args instead of just arg[0]	1
// TODO: much of preamble is common to both jitted_gpu_kernel and gpu_kernel	1
// TODO: Since there is some issues in gpu_kernel_multiple_outputs, we are	1
// TODO: remove the "buf_flat_size" process below and extend the buf bound	1
# TODO: Maybe add a separate section for the layernorm/dropout lstms	1
// TODO: This approach is a bit hacky as we assume that the second output is	1
//TODO: actually support int64_t index_t	1
# TODO: See if it's possible to use those directly.	2
// TODO: this should arguably support 3d as well	1
// TODO: Add actual patterns (like Conv-Relu).	1
// TODO: Deprecate this	1
# TODO (wanchaol): remove/merge this with ScriptLocalOptimizer once	1
// TODO: trick the optimizer for case where C % 4 == 0?	1
// TODO: not sure if we need to check for output size != 1, since we	1
// TODO: remove code duplication and unify code	2
// TODO (tugsuu): make sure this is optimized later	1
// TODO: changing at::ArrayRef<int> to at::ArrayRef<int64_t>?	1
# TODO: rccl_LIBRARIES should return fullpath to the library file,	1
TODO: We can optimize this to get rid of unneccesary transformation.	1
// TODO: The binary ops are copypasta.	1
// TODO: Ideally, this pass could run earlier, before inlining	1
# TODO write up issue, maybe fix	1
// TODO: code cleaning in parser so we don't need these.	1
# TODO: make this handle more cases	1
// TODO: It is possible to implement efficient batched orgqr for small tau (tau.size(-1) <= 32)	1
# TODO currently vgg19 doesn't work in the CI environment,	1
// TODO: Add hierarchical collapsible tree?	1
# TODO: Two weeks after this lands, combine these two overloads,	1
// TODO: this doesn't work with Scalar-Tensor ops! We should	1
// TODO: [T87180544] Implment softmax/log_softmax in metal shaders	1
/* FIXME: workaround for bug: https://github.com/pytorch/pytorch/issues/20342 */ \	1
// TODO: can batch_first be a wrapper around this function?	1
aliased_input = aliased_input_iv.toTensor();  // TODO: Can we avoid saving this tensor and incurring the refcount bump?	1
# TODO: Add type annotations	1
# TODO: https://msdata.visualstudio.com/Vienna/_workitems/edit/1408006	1
// TODO: Remove the last use of Device(const std::string& device_spec).	1
# TODO: better name	1
// TODO: Fix resize_as_. See pytorch/pytorch#11665.	1
// TODO: fix the constness of target	1
// TODO: merge this into populate_operands	1
// TODO: A useful internal assert would be to show that device_opt_ is null	1
# TODO: this only checks one input and one output, need to generalize to multiple	1
padding_mode: str = 'zeros',  # TODO: refine this type	3
/// TODO change this to const GraphT& g	1
// TODO Re-enable logging	1
# TODO: rename weight_is_statically_quantized to weight_is_int8_quantized	1
# FIXME: Ideally these functions should be methods on Type class, but we have a	1
// TODO: Set device to input	2
# TODO: xinyu, figure out why Nvidia do this?	1
TODO (mingzhe09088):	1
# TODO: handling of slice	2
// TODO: Calling add_context O(n) times has O(n^2) cost.  We can fix	1
// TODO optimize	1
// Run backward (TODO: make this async?).	1
input.ndimension() == 4 &&   // TODO: 5-D contiguous depthwise is not supported yet, need benchmarks	1
// TODO: not that it matters, but mark the right type here;	1
# TODO: I'm not sure if the clone here is necessary but it is safer	1
# TODO: maybe add a pass to cleanup bn modules?	1
// TODO: AutoNonVariableTypeMode should be removed in release 1.10.	1
// TODO: If we need TORCH_CHECK(available()) calls here as a sanity check, add it.	1
// TODO: Clean this up, Naoya added a mechanism we should be able to reuse.	1
# TODO: consider regex matching for test filtering.	1
// TODO: Support ComplexHalf accessor	1
// TODO: cut this over to HIP dispatch once we stop pretending that CUDA	1
// TODO: This attempts to keep the underlying memory alive by setting the base	1
// TODO: lower simple tuples ?	1
//! TODO: we also need to assert/check reduction axes and replace it with	1
# TODO: Update this once bfloat16 and float16 are better supported.	1
// TODO: check if all nodes have lowerings	1
// TODO: war_sync_ is only used for testing/validation purposes.	1
# TODO: make img_data a single example instead of a list	1
// TODO: make this point at hip_dispatch_ptr	1
# TODO: __reduce__	2
# FIXME: delete this	1
// TODO Kimish, we are allocating affine_quantized regardless of per channel or not.	1
// TODO: Better formatting.  Right-align this.	1
// TODO: why do we assume there's a single TV output?	1
# flaky test - TODO fix	1
// TODO: eventually, codegen these calculations and make them part of the	1
// TODO: change the types to vector<int64_t>	1
// TODO: remove or return launch parameters	1
// TODO: move this to a more shared place.	1
// TODO: If I put Undefined as entry 64 and then adjust the	1
// TODO: Rfactor a Welford	1
// TODO: Find way to expose alias info for opaque tensors.	1
{}  # backend_config_dict, TODO: point to README doc when it's ready	1
# TODO: use foreach API in optim._functional to do all the computation	1
# TODO: when https://github.com/pytorch/pytorch/issues/33782 is fixed	1
// TODO: this is only needed when dequantize_output_ == false but leave	2
# TODO: FIXME: RuntimeError: "bitwise_xor_cuda" not implemented for 'Half'	1
// TODO: Maybe distinguish between visible size and storage size.	1
# TODO: add more validations	1
# FIXME: mean does not support passing None to dim	1
// TODO: can we convert outputs_ to store indices?	1
# TODO: Validate ceil_mode semantics.	1
# TODO: update sample inputs with for_inplace_variant kwarg to support this test	2
// TODO optimize	1
// TODO exprs can have multiple accesses... we're returning the first but that	1
// TODO : omp parallelization	1
# FIXME: "prod_cpu" not implemented for 'Half'	1
// TODO: these groups should never merged into any other groups, but are	1
// TODO Licensing.	4
# TODO:  when python 2 support is dropped, change the signature to	1
// TODO: deprecate this in favor of aten::getelem	1
IntType::get()}, // TODO This type should be removed from the schema	1
// TODO: move as much of this into the constructors.	1
// TODO: T21635077 fix float-divide-by-zero undefined behavior	1
* TODO: Look into changing the threading semantics of Generators in ATen (e.g., making	1
// TODO: Enable view in parser by detecting non-alias view operation	1
// TODO: support?	1
// TODO: This is nuts!  There is no reason to let the default tensor type id	1
TODO: add dequant, if needed	1
// TODO: Explicitly test the 3 cases below	1
* TODO: https://github.com/pytorch/pytorch/issues/56482	1
// TODO: follow up with MKLDNN what the best way is	1
// TODO Find a better way to handle this.	1
// TODO: CUDA polling	1
// TODO: can use inplace ops?	1
# TODO: refactor this part to a function	1
// TODO: maybe this can go later in pipeline / directly in autodiff forward	1
// FIXME: this is a slow, simple implementation; need up/down sweep,	2
// TODO: We're going to get a lot of similar looking string literals	1
// TODO Replace by Reshape(), once wrappers are written	1
//!  TODO: Some expr simplifications could also be helpful	1
// TODO: assert requires_grad=False	1
# TODO: delete this!	1
raise AssertionError(f"TODO add support for type {repr(typ)}")	1
// TODO: consider if this should explicilty check for the file's existence or not to throw	1
// TODO: [T87350528] Fallback to shader kernels for 10.0 users	1
// TODO: Check if miopen has the functions above and unify	1
// TODO: support dynamic input by profiling it	1
except Exception as e:  # TODO: Catch only timeout exceptions	3
// TODO: Cleanup this.	1
// TODO: make it easier not to do O(k) iterations over the graph, where	1
// TODO: lookup by historic string key to start, then issue key	1
// TODO: add a test	3
# FIXME: improve load_tests() documentation here	1
# TODO: FIXME: RuntimeError: "max_elementwise_cuda" not implemented for 'ComplexFloat'	2
# TODO: Once we decide to break serialization FC, we can	1
# TODO: This method needs a refactor for clarity	1
// TODO: Only replay dispatch is really borrowed from TransformIter, we should	1
# TODO: remove dispatch section when porting TH CUDA to ATen	1
//!  (TODO: need structure better so we don't have to do this)	1
# TODO @ansley: add `Union` once landed	1
# TODO: Once we decide to break serialization FC, `storage` no longer needs to	2
(NOTE: Numbers below prior with missing parameter=update step, TODO to update)	1
# TODO: Possibly support variable-sized inputs.	1
// TODO: this class needs cleanup	1
# TODO this should be removed now that gpu support for quantization is being supported.	1
// TODO: handle dynamic shapes here.	1
// TODO: Check contiguous and dtype.	2
//    strong refcount           TODO: pack these into one word	1
// FIXME: this may not actually make any sense if we can efficiently move	1
// TODO: generalize logic	1
// FIXME: do these actually need to be zeros_like or can they be empty_like?	1
at::AutoDispatchBelowADInplaceOrView guard;  // TODO: remove	3
// TODO: Need to handle other Stmts / Exprs that read / write buffers.	1
// TODO: today, we put a single bailout template at the front to	1
// TODO: Potentially the following logic can be replaced by special logic in VariableType_x.cpp	1
// TODO Use list_element_from?	2
/// TODO: it's not in native_functions.yaml yet as it's not exposed to python	1
// TODO: Caffe2 Concat has an extra output. It should be only	1
// TODO: This should be something like TVMC2Frontend::supports(op);	1
# FIXME: mean reduces all dimensions when dim=[]	1
// TODO: maybe this should be a more shared location?	1
// TODO: This qint special case looks very suspicious...	1
// TODO We shouldn't use c10::impl stuff directly here. We should use the	1
# TODO currently onnx can't translate squeezenet :(	1
// TODO return generator object when torchscript supports RNG	1
// TODO Test use_count	1
# FIXME: min does not accept scalar inputs	1
// TODO: Brian Vaughan observed that we might be able to get this to work on	1
// TODO: consider printing more proper schema strings with defaults, optionals, etc.	1
// TODO: TensorArg check should start handle memory format	2
# TODO: complete the data type: bool, float16, byte, string	1
// TODO: Figure out why const-correctness doesn't work here	1
// TODO: review jiterating igamma and igammac if/when a persistent (across processes)	1
// TODO: Maybe eliminate required_size and just rely on next_pointer for bounds checking.	1
/// \todo FIXME: remove when \c constexpr becomes really \c constexpr	1
# hardcoded for now, TODO: expose the api to user,	1
// TODO: Remove???	1
# FIXME: nansum reduces all dimensions when dim=[]	1
// TODO: once modules support arbitrary ivalue attributes, we don't need this	1
// TODO: handle loop	1
// TODO: update this to support COO sparse layout	3
// TODO: temporarily disabled	1
// TODO: may need/want to initialize CUDA context here (refactor into nvrtc call)	1
# TODO: In principle, we track device information in our trace, so it	1
// TODO Disallow this and rather use std::unordered_map/set everywhere	1
*  - Pattern graph nodes cannot alias. TODO: the check not implemented yet.	1
# TODO: create dedicated columns	1
# TODO: remove the skip after these two operators schemas are fixed	1
# TODO: default exact_device to True	1
# TODO: review porting these to make_tensor	1
// TODO (possible optimization):	1
// TODO: Potentially more enforcements are necessary to avoid accidental	2
# FIXME add an override for JIT and revert 0. back to 0	1
// TODO: try more memory reuse algorithms and compare their memory efficiency.	1
# TODO: L1HingeEmbeddingCriterion	1
// FIXME: doesn't work for bias so we shouldn't quantize bias before	1
// TODO: we are sometimes emitting expressions like	1
// TODO: current RRef implementation does not tolerate failures	1
// TODO: more sane strategy	1
// TODO: handle channels last	1
TORCH_CHECK(std > 0.0, "normal_ expects std > 0.0, but found std=", std);  // TODO: dedupe	1
# TODO: add shape checks	2
# TODO: some signatures of std_mean do support out	1
// TODO: Handle non-contiguous Tensors.	1
// TODO: Consider readding Tensor and TensorList constructors here, when	1
// TODO - XXX - if any output is the same tensor multiple times, views	1
// TODO Add GPU support by writing a generic wrapper.	1
// TODO: consider unrolling CopyItems to make elemental types copy faster	1
// TODO: ListType case	1
# TODO: merge this to the case above?	1
// TODO: unify to C10_MOBILE. In theory this header could be used in OSS.	1
# FIXME: prod does not support passing None to dim	1
// TODO: This unwrapping code is ONLY used for TH bindings; once TH goes	1
// TODO: remove after empty Tensor serialization is forbidden	1
// TODO: Remove join()	1
# TODO: Once we have proper scoping, stop reimplementing chunk, delete this	1
// TODO: handle multiple kernels.	1
# TODO: FIXME: jiterator doesn't support non-tensor inputs	1
// TODO move to typeid.h (or codemod away) when TypeMeta et al	1
// TODO: can i haz an out version of the conv2d?	1
* TODO: move this to fbgemm after making transpose routine more general	1
// TODO clean up	1
# TODO properly map the exceptions in pybind (c10d/init.cpp)	1
// FIXME: try switch statement and explicitly handle cases	1
prim::unchecked_unwrap_optional, // TODO remove	1
// TODO Make quantize_tensor_per_channel_impl work for other datatypes too	1
// FIXME: it is actually possible for us to handle padding, figure	1
// TODO: should be b.transpose(1, 0)?	2
// TODO: add the max/min clip	1
// TODO this should go in the global Python CU	1
// TODO: return iterator	2
# TODO: Add calibration for the sparsity	2
# TODO: sure looks like we unconditionally initialize the context here	1
# TODO: RuntimeError: cholesky_inverse does not support automatic differentiation for outputs	1
// TODO: channels last 3d	1
# TODO: Consolidate `i0e` with sample_inputs_unary when `make_tensor`,	1
# TODO: when FL is migrated from full-jit to lite trainer, remove '__ROOT__'	1
# TODO: Delete this line once https://github.com/pytorch/pytorch/pull/55889 lands	1
# hacky check for collections.namedtuple, TODO improve this	1
// TODO move this to c10 namespace	1
# TODO: Handle dilation	1
# TODO: __name__ not set for submodules in recursive script	1
// TODO: type is unknown until a user starts to fill data;	1
# TODO: type dim as BroadcastingList when	1
# TODO: remove these flags when	1
// TODO: verify the same thing in miopen	1
# TODO: Once we decide to break serialization FC, this case	2
// TODO: only handling conv2d at this moment, expand this to convXd	1
# TODO (wanchaol): remove this once we added TorchScript	1
// TODO: add %dtype after when https://github.com/pytorch/pytorch/issues/34351	1
// TODO Can be replaced with packB->getOutputChannels() when update pre-pack	1
# FIXME: mean does not support passing keepdim without passing dim	1
# TODO: Add Gumbel-Laplace KL Divergence	1
# TODO: assertions could be expanded with the error messages	1
// TODO: Instead of cudaStreamSynchronize it is possible to add Stream	1
# TODO: These are deprecated, maybe we shouldn't type hint them	1
# TODO: create a dedicated SelfArgument type for 'self'?	1
// TODO: total hack. Switch to numel when it is available.	1
//!  TODO: unify the segmented and un-segmented code-path	2
# TODO: Need to handle collisions with argument names at some point	2
// TODO: split out	1
// TODO: Look up cached contained types. this is kind of tricky	1
# TODO: output dtype in the config and  parse it back from the str	1
# TODO: --op_registration_whitelist will be removed when all call-sites	1
// TODO: scalar-tensor ops should be canonicalized	1
// TODO: Once copy_ is fully migrated to use dispatcher, handle named	1
// TODO: replace Half by BFloat16, after BFloat16 is supported by Nvidia	2
// TODO: support channels last in sum	1
// FIXME: really, overlapping writes should be illegal/an error in Torch	1
// TODO: remove this code path once Variable and Tensor are merged in Python	1
* FIXME: Can we specialize elementwise_kernel and launch_kernel in Loops.cuh	1
//TODO: [ROCm] Need to remove this after CUDA->HIP mapping is updated.	1
# FIXME: prod reduces all dimensions when dim=[]	2
// TODO: use tensor.index() after improving perf	1
# TODO: the name should be weight is int8 quantized	1
// TODO: Workaround since MIOpen does not support NHWC bias	2
// TODO: jiterate kaiser window and make them only available when not jiterating	1
// TODO: handle strides?	1
// TODO: fix it after the storage sharing is merged	1
return f.func.name.name.base  # TODO: should be str(f.func.name.name)?	1
// TODO: enable input fillers	1
// TODO: fix this when windows can correctly capture variables in nested lambda	2
# FIXME: Once 3.7 is the minimum version, type annotate `other` per PEP 563	1
// TODO: _batch_norm_impl_index_backward is only used in JIT. cudnn NHWC	1
// TODO Replace by Reshape(), once wrappers are written	1
// TODO: use nccl reduce	1
// TODO: Make HIPIFY understand CUDART_VERSION macro	1
// TODO: contiguous is called for further jit optimizations.	3
# TODO: implement add_param_group	1
// TODO: Re-audit this; it used to be an indexSelect directly into r_values	1
// TODO: think more about TensorExpr alias correctness	1
// TODO: This is a known bug that it's not yet implemented and the allocation is failing.	1
// TODO: if we need to, we can also enable this path for quantized tensor	1
// TODO: Rewrite this using dispatchKeyToTensorOptions	1
# TODO: hip_hcc has an interface include flag "-hc" which is only	1
// TODO: Maybe skip this for fixed-size outputs?	1
# TODO: There is a bug in rocblas's & rocfft's cmake files that exports the wrong targets name in ${rocblas_LIBRARIES}	1
// TODO: Support more dtypes.	1
// TODO: implement equality	1
# TODO: reference function	1
# FIXME: does not support passing keepdim without dim	2
// TODO: first operand for pow can be Tensor / Scalar	1
// TODO: this could be insertReferencedAttr to be more clear,	1
// TODO: canonicalize as aten::dim ?	1
// TODO: There is no need to branch with every element	1
// TODO: packing + quantization in same block.	1
// TODO: cache these result from the forward pass	1
// TODO: channels last kernel can be made faster.	1
// TODO: implement _out variant avoiding copy and using already allocated storage directly	5
# TODO: fix torch.zeros(sizes, grad.options()) before enabling select, topk, kthvalue	1
# TODO: current patterns are the ones after fusion, we will want to expose fusion	1
// TODO: check for overflow	1
// TODO: we need to figure out what are supported input scalar	1
// TODO: When adding failure retries and timeout, this fork needs to be	1
# TODO: Update the comment reference to the correct location	1
# TODO: Remove this dependency	1
// TODO: check type and return the right flag	1
# TODO remove Dropout special after codebase stable	1
// TODO: not sure if optimizer is able to compile two levels of	1
// TODO: implement negative step	1
/* TODO: TORCH_CHECK just have 2 args: condition and message */	1
// TODO: replace "PredictorParameters" with the constant in OSS bbp	1
# Stores underlying RecordFunction as a tensor. TODO: move to custom	1
// TODO:  Write a more refined GRU heuristic.	1
.target_device(options.device())  // TODO: this shouldn't be necessary if it came from options	1
# TODO (alband) Why this is not automatically broadcasted? (had to add the repeat)	1
// TODO: incorporate center distance normalization.	1
# FIXME: does not support dim=None	2
// TODO: avoid extra copy by directly feed initializers to backend blobs	1
// TODO: limit it only to amp related node;	2
// FIXME: crashes if exception type is not RuntimeError	1
# TODO: This discrepancy isn't required; we could also generated	1
// TODO: this doesn't work with Scalar-Tensor ops! We should	2
# TODO: This implies that ellipses is valid syntax for allocating	1
// TODO: we should support any type, not just float	1
// TODO: I think we should do this but	1
# TODO: refactor this part to return WEIGHT_INDEX_DICT and BIAS_INDEX_DICT	1
// FIXME The two resizes below zero out the vectors, which is not needed.	1
# TODO: This is a potential source of accuracy drop.	1
# TODO: review with var_mean tests in test_autograd.py	1
// TODO reuse col_offsets_with_zero_pt_s8acc32_ref in fbgemm	1
# TODO: Add an example to use such a wrapper.	1
# TODO: Add Gamma-Laplace KL Divergence	1
// TODO: elaborate in this comment on the structure of math.cuh	1
// TODO: We currently expect contiguous memory layout.	1
// TODO: Maybe 260 KB is a bit small...	1
# TODO: Call the torch.ao.utils.convert in here	1
# TODO: Handle this case better.  TorchScript ranges are in bytes,	1
// TODO: remove output dequantization once NNC supports quantized outputs.	1
// TODO: Maybe more user friendly to report where the expected size	1
// TODO: restore the above, see https://github.com/pytorch/pytorch/issues/64709	2
# TODO: after having proper ways to map Python strings to ATen Enum, move	1
// TODO: rotated_rect_intersection_pts is a replacement function for	1
// TODO: remove when possible, since it just slows down	1
# TODO: We should really error in this case, but its bc-breaking so	2
// TODO: check x for consistency with input_size?	1
// TODO: put this in the correct device???	1
// TODO: Make this a real argument	1
# TODO: clone storage aliasing	1
// TODO benchmark to decide whether to remove this special case	1
// TODO: tune this with cache size detection code. Changing to 32 helps on some	1
#                       TODO: bfloat16 support.	1
// TODO: add requires_grad once we decide on semantics for sharing data.	1
//! TODO: we should probably change shared_ptr to unique_ptr, as we want to	1
TODO: when scale != 1 is introduced then use:	1
# TODO blowtorch	1
// TODO: switch to PyTorch dtype as it's closer to truth.	1
// TODO: Make rpc_op(..) support taking kwargs,	1
# TODO: We will save the original weight and bias, because the unpacking is not yet there.	1
// TODO: Add documentation.	1
// TODO: even though both scopes are conditional, we can merge accesses if	1
# FIXME: torchscript: torch.zeros(sizes, grad.options())	3
// TODO: for (1), multiple -1 may conflict each other. Consider use	1
// TODO: add parallelization once fbgemmGroupwiseConv supports multi-threading	1
// TODO: consider caching the heuristics value so tryMerge doesn't have to be	1
// TODO: hadd_pd() & hsub_pd() may have scope for improvement.	2
# TODO Add some asserts about input type	1
// FIXME: This probably should be called WorkGloo since the work is executed in sync mode	1
# TODO: can we use JIT here to reduce python overhead?	1
// TODO At some point this should probably be done, including tricks	1
# TODO: PartialLinear - maybe in sparse?	1
# FIXME: figure out a better way when we support sparse tensors in jit	1
// TODO: Unify 2d and 3d when ChannelsLast3d is ready.	1
# TODO: @kefeilu: also incorporates a validator to do inference (and optionally)	1
# TODO: These functions are not used outside the `fuse_modules.py`	1
// TODO: perhaps it would be nice to have int128, a signed 128-bit type?	1
# TODO: Do this in translate instead	1
// TODO: expand support to wire non-constant inputs, this is currently	1
# TODO: Move orelse to the body after calling ConditionalExceptionWrapper.	1
// TODO: hacky way of deciding the groups	1
// TODO: add a note explaining the design decisions	1
// TODO: update this when codegen can output scalar	1
# TODO: merge with get_static_quant_module_class	1
// TODO: We could squeeze some perf by calling at::cuda::mul_out here instead, to bypass the dispatcher.	1
# TODO: remove this as onnx opset 11 spec allows negative axes	2
# FIXME: logical_or does not accept scalar inputs	1
// TODO: convert to c10::optional	1
// TODO: I'm assuming we have stride information in `graph->toString`	1
# TODO: create an internal helper function and extract the duplicate code in FP16_compress and BF16_compress.	1
# TODO: we should probably add them in	1
// TODO: ideally we should be strict here and return nullopt if the dtype is	1
// For cases with scripting, TODO: Add logic to handle NoneType outputs	1
// this should be removed. TODO (kimishpatel)	1
// TODO Update quantize_tensor_per_channel_impl implementation to follow	1
//TODO use _assert_fail, because assert is disabled in non-debug builds	1
# TODO: delete the original weights	2
// TODO: replace conv1d with conv2d ?	1
//!  TODO: refactor this similar to isConsumerOf	1
// TODO: If caching inputs would require persistence we are sending it to the	1
// FIXME: by doing this only on the inputs, we only capture graph inputs and	1
# TODO: is_reference option for conv module	1
DecorateInfo(unittest.skip("FIXME: Jiterator does not support complex outs!"),	1
# TODO: consider adding torch.unravel_index	1
* TODO: make this API also take "offsets" rather than "lengths" to match the	2
this->Size = this->Capacity = 0; // FIXME: Setting Capacity to 0 is suspect.	1
// TODO: Add the assert after all uninitialized states are eliminated	1
# TODO: Update comment below since it is out of date.	1
// TODO: it seems like sparse_dim == 0 could be supported even if self.dim() >	1
// TODO: after we add Get<Tensor>(DeviceType)	1
# TODO: update this when promote_types supports bfloat16 and/or	1
// TODO: figure out how to do chaining	1
# TODO finish disentangling control flow so we don't do in-projections when statics are passed	2
// TODO: run root tasks inline in inference mode	1
// TODO: Remove this else. Or add assert	1
rpc_backend_options (rpc): configurations/options for the rpc TODO: fix	1
# TODO: Determine if the other cases need to be fixed as well	1
// TODO: Maybe show key?	1
// TODO: Deduplicate from compute_at.cpp	1
// TODO: Maybe TensorAccessor can be used here?	1
# TODO: properly handle case when u is tuple instead of only taking first element	1
// TODO Direct implementation might be faster	1
# FIXME: document this and move it to test_serialization	1
# TODO (zaf): Inherit from `quantized.Linear` (T83294430)	1
// TODO: Is it worth it to have a contiguous call or maybe we should go with	1
// TODO: Revisit this once we decide on how dependencies analysis should look	1
// TODO: replace __syncthreads with __threadfence for alias ops	1
// TODO: this is not a legit assumption? Can't we run with	1
# TODO: rename this to supports_bwgrad_bwgrad to be consistent with below	1
// TODO: we are only checking output 0. This means that our current check for	1
// TODO: do HIP	1
// TODO: revisit this later to use batch_iterator_with_broadcasting in triangular_solve	1
base_name = f.func.name.name.base  # TODO: should be str(f.func.name.name)?	1
//     loop variable. TODO: Remove this constraint.	1
// TODO: move to this an internal IR.	1
// TODO: put this into ir_cloner instead	1
// TODO: are these a single call to cublas batched matmul?	4
// TODO: dont require # of dimensions of tensors set ?	1
// FIXME: it's vA < vB because the sorting order for V (aka	1
# TODO: maybe the logic to search for all variants is no longer necessary?	1
# TODO: Make this an enum.	1
# FIXME: remove this by updating test suites using it	2
// TODO: investigate parallelization of the accumulate kernel.	1
ret = input  # TODO: remove when jit supports exception flow	3
# TODO: What exactly is the semantics of the 'dispatch' field?	1
// TODO: use guards to avoid leaking	1
// TODO (@zasdfgbnm): this function assume trivial 1d and no dynamic casting	1
// TODO: This will only be useful if we write a backend fallback that plumbs dispatch keys (currently there are none)	1
// TODO: refactor all current uses of this function to the Opt one	1
// TODO: Move this function to MetalContext	1
// TODO: fast exp	1
// TODO: Allow individual functions to specify non-default translations:	1
//    TODO: make sure there's nothing wrong with grouping of nodes that	1
# TODO: Why skip this? Because @torch.jit._overload_method will	1
// TODO: enable other min/max variants, operators that can be both	1
// TODO: this should be handled during dispatch, but that's missing...	1
// FIXME: we don't actually handle a dynamic padL > 0	1
//!  TODO: remove inline versions of this as much as possible	1
# FIXME: figure out the flaky -1024 anti-leaks on windows. See #8044	1
# TODO: handle the other Ju	1
// TODO: specialize fbgemm::Quantize for a single vector and make it	1
// TODO: need to call GenerateClone sometimes? Or else return LowerBuiltIn() directly	1
// TODO: runOnGraph takes const ref?	1
// TODO: don't strictly need to reset write cache, evaluate on models	2
// TODO: move this to Buf printer	1
// TODO: 1. Add assert in the dist engine to ensure no GPU NodeTasks during	1
# TODO: consider more complicated noncontiguity schemes	1
// TODO: Handle all other cases here.	1
// TODO: Decouple and improve error handling and messages.	4
// TODO: free GIL - but remember to reacquire it when an exception is thrown	1
// TODO: caffe2::PThreadPool only provides a data-parallel API.	2
# TODO: write a non-layer checker and log it	1
// TODO: Update alias key precedence after we add new alias keys AutogradDispatchCPUOrCUDA .	1
// TODO: Deprecate me	1
// TODO: refactor bindFusionInputs to better support this	1
points for compilation (TODO add a link when the rules are published).	1
// TODO: Use name-hint of the producer indices instead of 'idx'	1
//!    TODO: make sure there's nothing wrong with segmentation on nodes that	1
// TODO: can be unified with at::from_blob when Tensor is merged and string	1
// TODO: trick the optimizer for case where C == 4?	1
# TODO: discuss a unified TorchScript-friendly API for autocast	2
* // TODO modify docs	1
- Fix various TODO comments in this file and the JS.	1
// TODO: Activation transpose before and after the kernel can be removed if we	1
// TODO we need to refactor graph APIs (e.g., addInputs)	1
# TODO: What follows is a reference implementation of a masked sum	1
# TODO: it's possible that the following is confusing:	1
// TODO: remove following two after at::kDouble and its friends are TypeMeta's.	1
// TODO: This is annoying, having to put the cudnnTensorDescriptor_t	1
// TODO: Make this accept options instead of dispatch key	1
# TODO: this is not handling non-tensor tuple args (for example,	1
// TODO: investigate parallelization of the accumulate kernel. Unlike the non-accumulate case,	1
# TODO: handle objects with deeper nested tensors	1
/// TODO: Eliminate this function as much as possible, as it can be expressed	1
// TODO This should be deprecated in favor of linalg_matrix_exp_differential	2
// TODO: When lint fails, give better indication about which	1
# TODO: need to specify this is side-effectful somehow	1
// TODO: Use name-hint of the producer instead of "temp"	1
// TODO: can non_c10_complex go through the other path?  Need to verify.	2
inline void Record(Event* ev, const char*&) const { /* TODO */	1
// TODO: we could attempt to follow the GetAttr chain and	1
// TODO: Use itensor after 0-dim is supported. Now use CPU tensor.	3
// TODO: Issue #20497	2
// XXX TODO: revisit the alternatives	1
# TODO: replace torch.subtract/divide/square/maximum with	1
# TODO Movie this to C++ once the jit has better support for torch.Size.	1
# TODO: some signatures of var_mean do support out	1
// TODO: failure in buildShapeExpressions should not break fusion execution,	1
# TODO: maybe create a PythonTensorOptionsArgument?	1
# TODO: this is looking into how the value is used in the future	1
// TODO: Make this actually return something that's "user friendly".	1
// TODO: empirically, on OS X this assert appears to be untrue	1
// TODO: figure out if we can narrow gO and save some compute,	1
// Trigger tests for D25440771. TODO: Remove this line any time you want.	1
# TODO: support get_autocast_gpu/cpu_dtype	1
// TODO: Do we have a ScalarOrTensor type?  Would such a thing exist?	1
// FIXME: _fft does not support complex_output=false with inverse=false	1
# TODO: There's an issue here with FC. It might be impossible to	1
// TODO Making operator!= noexcept if operator== is noexcept doesn't work with	1
// TODO: This file is still in the caffe2 namespace, despite living	1
// TODO Handle other composite types, such as vector<...>	1
# TODO: expose other parameters in the future.	1
# TODO: Conv2dMap	1
TODO: Need a clean way of loading the state of the "preapred" module	1
// TODO: Cache indices in forward pass to re-use in backward	1
# TODO: Why do I have to call this grad?!	1
// TODO: this is not exposed here, I need to remove that before inserting	2
// TODO: This code can path can be removed if #61309 is resolved	1
// TODO this should come from cmake	1
// TODO: We might need to use nodes_map instead of value_map. Otherwise, we	1
// TODO: pass tracer and counters through ExecutorHelper	1
// TODO: uncomment when we properly support pow	1
// TODO: share more logic with tensorexpr_fuser ?	1
# FIXME: keepdim parameter is ignored when dim=None	2
# TODO: Transform at load time to share weights with CPU model.	3
// TODO: Reuse convertToDotString once convertToDotString can work	1
# TODO: Add support for multiple parametrizations for the same weight	1
// TODO: RRef internal messages are not yet idempotent	1
// TODO: verify that the CUDA handles copy from device to device correctly	1
// TODO: Add an IR verifier check to detect invalidly compressed buffers.	1
# TODO: Once we decide to break serialization FC, no longer	2
// TODO: Handle nested dictionaries.	1
# TODO: more robust handling of recognizing ignore context manager	1
// TODO: Use bit_cast once C++20 becomes available.	1
// FIXME choose desired unroll level	1
# TODO from zdevito:	1
// TODO: use a dedicated bind-var to make sure v is not evalualted multiple	1
# FIXME: torchscript: int - bool	1
// TODO: may be more efficient to place after the first non-computeAt	1
/// TODO: The consistency check here is inconsistent with StreamGuard's	1
assert len(schema.keyword_values) == 0, "TODO the logic for operand(i) is broken if there are kw values"	1
// TODO thread_local might fix this	1
// TODO Union types are not supported on embedded runtime, and we need to	1
// TODO:: check restrictions for inputs; outputs not used elsewhere	6
// TODO: consider adding a 32bit indexed kernel for improved performance	1
using Stack = torch::jit::Stack; // TODO Instead of this, move torch::jit::Stack to the c10 namespace.	2
// TODO: Would be great if we didn't need this, but we have nice functionality	1
# TODO: clone indices in sparse tensor ctor.	1
name = f.func.arguments.out[0].name  # TODO: old codegen behavior - should fix	1
// TODO: add memory types.	1
// TODO: delete this and update accessor for value_map(_)	1
# TODO: figure out how to return torch.return_types.histogramdd	1
// TODO: turn contiguous_ into an enum CONTIGUOUS, NONCONTIGUOUS,	1
// FIXME: TestJit.test_ge_optimized fails this assertion.	1
// TODO: Make this more efficient	1
// TODO: consider using C++ stack trace	1
// TODO: When handling fused ops d = a + b + c, the correct	1
// TODO: Support only 9 args once the old signature has been removed.	1
# TODO: return this to the caller	1
# TODO : Fix these discrepancies	1
// TODO: here, we create a Storage	1
// TODO: Perhaps we should use cloneFrom now, as it seems unlikely	2
// TODO remove these switches once interface call is rolled out.	1
# TODO: check schema	1
// TODO: merge this with at::empty after Tensor is merged	1
attr.isTensor()) { // TODO: Handle float/int	1
TODO: refactor so we don't set inputs onto the subgraphs	1
// TODO: consider using C++ stack trace	1
# TODO: See if we can remove this in the future	1
// TODO: this should be replaced by RAII wrappers.	1
// TODO: Maybe track by dtype as well.	1
raise unittest.SkipTest('TODO: Memory availability checks for XLA?')	1
# TODO: Determine whether this can be removed after type inference.	1
// TODO - Implement rfactor	1
# TODO: Uniform-Laplace KL Divergence	1
// TODO: Do we need this, or can we get it from getMatcherGraph?	1
# TODO: This was historically used to help some JIT interop code	1
# fbgemm (FIXME: this assumes fbgemm is used only for NHWC and im2col	1
// TODO: Consolidate this file with util.h	1
# TODO: Update all quantization tests to use this decorator.	1
// TODO : `unary_jitted_gpu_kernel` for cleaner UX.	1
{}); // TODO: https://github.com/pytorch/pytorch/issues/55757	1
* TODO: This will only be useful if we write a backend fallback that plumbs dispatch keys (currently there are none)	1
// TODO: int -> int64_t	1
ss << "{TODO implement Node::shape}";	1
tmp_y[i] = _cvtss_sh(y[i], 0); // TODO: vectorize	1
// FIXME: In case forward has multi outputs, we only support one requires grad	1
# TODO: support dynamic quant	1
# FIXME: heaviside does not accept scalar inputs	1
TODO: This is fragile, whether output is quantized should not depend on `is_reference` since	1
# FIXME: cannot specify keepdim without dim	2
// TODO: arange doesn't have complex support	2
1); // TODO: expose as argument?	1
#                       TODO: optional start/end attribute.	1
// TODO: gcc vectorization	2
// TODO: Find way to expose alias info for opaque tensors.	1
// TODO: MemoryFormat is not implemented in this way	1
# TODO: Maybe embed the enforced zero_point in the `torch.iinfo`.	1
// TODO: remove this comment after enabling autograd support for CSR tensor	1
// TODO: Reduce this extra TensorIterator construction for Reduction::Mean & Sum.	1
// TODO: (@pavithran) size is overloaded with int[] and Tensor	1
seed=random.randint(0, 100000),  # TODO: dropout seed	1
// TODO: unlock GIL when contacting the manager	1
// TODO: call into `convertOutputToCorrectStrides`. Currently this causes a	1
// TODO Needs to be fixed this to work in all cases	1
// TODO: We should fix this when we have some time.	1
// TODO: may be more efficient to place before the first non-computeAt	1
// TODO:  try_get_weight_buf returns a Tensor, but _cudnn_rnn below takes a c10::optional<Tensor>	1
# TODO: enforce that functions in fp32_to_int8_fun_mapping must both be	1
// TODO these arrays are potentially of the different types, use function	1
# TODO: there should be a more principled way of doing this.	1
# TODO: When `len(inputs) == 1` and all inputs are on `destination`, just	1
// TODO: make this cheaper by computing hash of fdata.	1
// TODO: Resize and copy should be avoided with	1
// TODO: remove when the function in math are changed to use vector<int64_t>	1
// TODO: replace input.svd with linalg_svd	1
// TODO: Ideally AutoNonVariableTypeMode in this file should be changed to	1
// TODO: move to TensorMath.cpp	1
# TODO: Should we do this even for non-contiguous tensors?	1
// TODO: we can actually support it by adding an extra inputs to the	1
// TODO: To support all features of MemoryFormat::Preserve we need to add	1
// TODO need guards on init returning none	1
# TODO: move all tri/tril/triu testing to tensor creation op test suite and remove	1
// TODO: Type Inference does not propagate Shape Information	1
# TODO: it's not used, so actually we can skip quantization	1
// TODO: support input tensor with dynamic shape (PR #54982)	1
// Brute-force algorithm. TODO: Come up with something better.	1
# TODO: Need to add options to qconfig to avoid the calibration.	2
// TODO: revisit this (do we need to consider float64 types?)	1
// TODO: call UDF if it exists	1
// TODO: refactor codegenOutputQuery into its own file	1
# TODO: we should probably handle a few additional errors,	1
// TODO add other op signatures.	1
// TODO: there should be a shorter way to spell this	1
// TODO: The following don't work on Windows. Specifically, sigaction, waitid	1
// TODO: This is a very naive implementation with a single mutex. We can do the	1
# TODO: we may want to try to remove the special case here	1
// TODO: add LSTM function to composite operations	1
# TODO: make sure virtual devices such as 'cpu1' and 'cpu4' are supported.	1
// TODO: remove later	2
// TODO This backward pass uses a very complext expression to compute (diff	1
// TODO: separate passes into different file;	1
# TODO: include the configuration in backend_config_dict	1
# TODO: Once we decide to break serialization FC, we can	1
# TODO: only to keep it byte-for-byte compatible with the old codegen, should remove.	2
// TODO: Use something less heavy duty than a big honking mutex	1
// TODO: Is there a way to py::cast that doesn't raise an exception on	1
return NamedCType(binds, MutRefCType(BaseCType(tensorT)))  # TODO: fix this discrepancy	1
// TODO: lift the restriction of not fusing producer containing reduction when	1
// FIXME: remove this check once cub sort supports bool	1
// TODO: should we set this as human-readable time instead of unixtime?	1
for (int i = 0; i < N; ++i) { // TODO: multithreading	5
schema.overload_name().empty(), // @TODO: is this check correct?	1
// FIXME: _fft_r2c doesn't support native r2c IFFT	1
// TODO: class types cannot be redefined because we have no way right now	1
// TODO: static_assert that a templated function exists, and throw a friendly	2
# TODO: remove them when users are ready to take a hard dependency on PyTorch 1.	1
// TODO: move this somewhere more generally useful	1
// TODO @wconstab refactor to use ModuleValue::asTuple instead of new API	1
// TODO: Make this call the TensorOptions version, maybe?	2
// FIXME: this doesn't actually unroll; clang has per-loop unroll	1
// TODO: C++17 has the fileystem header, which may replace these	1
// TODO: replace Half by BFloat16, after BFloat16 is supported by Nvidia	1
# TODO: Kill this when we eventually remove it!	3
# TODO: make sharding spec a ChunkShardingSpec by inferring from the metadata list.	2
// TODO: Only remaining use of this is in index compute, remove use from there,	1
# TODO: try to recover the location of else:? Python doesn't give us useful	1
// TODO: The question about how to handle negative orders when the input	1
// TODO: going to need to change this if we want nested functionalize() transforms.	1
// TODO: change the number of microprocessors	1
// TODO: enable slice, shape inference is not implemented for this op yet	1
// TODO: Handle arg promotion.	4
// TODO: Stop calling PyObject_HasAttrString() in a loop on our read loop	1
# TODO: deprecate this	1
// TODO: #ProfileIValue List should update this	1
// TODO: Error if an operator is def'ed multiple times.  Right now we just	1
//!  TODO: Space efficiency of this class will be important,	1
// TODO: eliminate const_cast	1
// FIXME: We'd have to find some other trick with Thrust to perform a	1
# TODO: See if we can remove this in the future if we are	1
# TODO: To cover more problematic cases, replace stride = 0 check with	1
# TODO: update this when isclose is implemented for CPU float16	1
// TODO: enable 3d batchnorm.	1
// TODO: This qint special case looks very suspicious...	1
// TODO dim == 4 case will be enabled once it is fully tested	2
// TODO: not supporting casting to outputs is only really necessary for arg{min,max}	1
// TODO: uncomment once we can handle rand+broadcasts	1
// TODO: we need to check input type when we handle `to()`	1
# TODO: handle other tuple subclasses more generically	1
// TODO: is it worth optimizing this loop via padding in C?	1
# TODO: for _shared_model, do only NCCLReduce	1
// TODO: checking this is not free, so we should stop if this keeps	1
// TODO: This is better represented as an OrderedDict, but alas it is not yet	1
# FIXME: logical_and does not accept scalar inputs	1
# TODO: Once we decide to break serialization FC, no longer	1
// FIXME: We need to call it here since Future completion requires all	1
# TODO: add args/kwargs for passing to assertEqual (e.g. rtol, atol)	1
# TODO: https://github.com/pytorch/pytorch/issues/53023	1
// TODO: Write in more idiomatic C++17	1
// TODO: come up with a more user friendly interface	1
// FIXME : code duplication with	3
auto dx = at::empty(input.sizes(), input.options()); // TODO: more compact way of saying this	1
# FIXME: nansum does not support passing keepdim without passing dim	1
# TODO: enable later	1
// TODO: This is bad; this test should apply universally	1
# TODO: I think this means structured won't work with method	1
// TODO: pass pre-resized output?	1
// TODO: factor out isCallFunc	1
# TODO: Can we improve this error message to point out the gaps?	1
// TODO: implement MAGMA-based path using magma_zgeqrf_expert_batched	1
// TODO: clarify hasSideEffects, isNondeterministic	1
// TODO: use CUDAGuard here instead of context and employ explicit sync	1
# TODO: not sure if typing supports recursive data types	1
// TODO: It's possible this is still triggering	1
// TODO: add support for the following fusible operators.	1
// TODO: Use shape information from weight tensor	1
# FIXME: dim=[] reduces all dimensions	3
// TODO: support the case where kernel allocates output tensors dynamically.	1
// TODO: support int64_t dims in ideep::tensor to avoid extra conversion	2
// TODO: kernel implementation could stride on spatial dimension. We probably	1
// TODO: hacky way of inferring the groups number for grouped Conv3D	1
// TODO: This may be a common operation, could be worth making a utility	1
# TODO: Support automatic reshape	1
# TODO If this codepath becomes popular, it may be worth	1
# TODO: Update to provide the libraries and paths for linking npymath lib.	1
// TODO: pass block_size here;	1
// TODO: retry on collision	1
// TODO: change to a proper error report	2
// TODO: Improved heuristic on when to coalesce or remove need to coalesce	1
// TODO: old fuser is not maintained internally, somewhere it is being turned on	1
// TODO: Add human-readable sizes?	1
// TODO: Should we revert to the original timeout at the end of the call?	1
# TODO: fix lint	1
// TODO: only check `!impl_->requires_grad()` after Variable and Tensor are	2
// TODO: update cast_op signature to take dynamic context flags	1
// TODO: this should be merged with the storage flattener.	1
# TODO: we don't have _concrete_type set after load(), and in general we lose constant information.	1
# TODO: Uncomment when negative weights is supported.	1
//FIXME - this are defined in Loops.cuh, but including Loops.cuh here would lead to circular includes Loops.cuh -> CUDALoops.cuh -> jit_utils.h -> Loops.cuh	1
// TODO This probably should be using at::native::make_reduction	1
// TODO: Fill in output sizes	1
// TODO: dedup this part with code in quantizeTensors	1
// TODO: make IR nodes extensible.	1
# TODO: remove these special cases, ArrayRef fallthrough works fine	1
// TODO: Stop using legacyExtractDispatchKey here (probably need to build	1
# TODO: We need to change this to rpc.remote, and make it async (see the else branch below).	1
# TODO: This code can path can be removed if #61309 is resolved	1
// TODO: Remove PYTORCH_MIOPEN_SUGGEST_NHWC once ROCm officially supports NHWC in MIOpen	1
# TODO: maybe rename this to MatchInputNode	1
// TODO: split it into two kernels to make it more similar to exact	2
// TODO: Remove this function when at::native::empty() is modified to accept a	1
// TODO: unhook this	1
// TODO: insert specialized cases (e.g. depthwise convolutions, the direct	1
// TODO: check consistency, e.g.: code version, input shape and compiled	1
// TODO: qscheme	1
* TODO Instead of doing it this way, we should only have pure-jit ops in	1
// TODO: There are several places that recurse over IValue. This is fragile.	1
// TODO: Raise if not all output values are visible in input geometry.	1
# TODO: this can be simplified after https://github.com/pytorch/pytorch/issues/69316 is fixed	1
// TODO: fix this when windows can correctly capture variables in nested lambda	1
// TODO: is this necessary?  We used to treat nullptr-vs-not in IntList differently	1
# TODO: can we emit the union of these? What are the implications on TorchScript	1
// TODO: replace this with a real overload_name when FunctionSchema supports	1
// TODO - Implement cache_access	1
// TODO: Any chance to make this cleaner?	1
// TODO: uncomment the following when svd is deprecated not only in docs	1
// TODO: XZP kernels won't be supporting per channel quantization.	1
# TODO: (@krshrimali), add error_inputs_func once https://github.com/pytorch/pytorch/pull/67354 is merged	1
// TODO: Rename to register	5
// TODO: add a mutex to make it thread safe.	1
// TODO: refactor findObserverName to take Node* as input	1
// TODO: not currently using these gradients, investigate t16675365	1
// TODO: we should use wrapped pg_'s timeout here, but C++ ProcessGroup API	1
// TODO: does this works?	1
// TODO: This code path is not ideal as we are "lying" to the caller in	1
// TODO: consider using trash can	1
# TODO: figure out one liner to .clone() and set requires_grad	1
# TODO: why this needs to be special case?	1
# TODO: fixme	1
// TODO: hack to make `test_lstm_gates_permutations_cuda`	1
// TODO: we need to figure out how to profile calls to custom functions	1
# TODO: enable inplace in aten exporting mode.	2
// TODO: should we consider adding support for NoneType;	1
# TODO: fix along with var_mean autograd tests	1
# FIXME: maximum does not accept scalar inputs	2
// TODO: check multiple uses ?	1
// TODO: Check adding iteration domain unrolling	1
// TODO: update and add a usage example after https://github.com/pytorch/pytorch/pull/58092 lands.	1
# TODO: Properly handle aliasing caused by get_attr. For now,	1
// TODO Is there some way to implement this?	1
# TODO: improve the handling of complex tensors here	1
# TODO: Support this by adding trailing 1 dims.	1
// TODO: Turn this into an honest to goodness class. I briefly attempted to do	1
// TODO: do MLC	1
// TODO: We do go through different code path, should investigate whether this	1
// TODO: The logic for cast_outputs will need to be handled by the	1
// TODO Do schema inference without relying on WrapFunctionIntoFunctor	4
// TODO: better TensorOptions argument passing(e.g. default argument)	1
# TODO: Get rid of dynamic_type, after getting tools/autograd	1
# TODO: In principle, we could provide more structured version/config	1
// TODO: could this be a reference and not allocated on	1
# TODO (refactor) this is duplicated, maybe have a helper function	1
// FIXME: remove magic > 0 after we ensure no models were serialized with -1 defaults.	2
// TODO - expensive (>1ms) - cache these.	1
# TODO: Uggggh, parsing the schema string here, really???	1
# TODO: Add GPU support	1
// TODO maybe avoid call to vec	1
//FIXME I didn't find how complex -> real conversion is done in eager	1
// TODO - come up with better message	1
// FIXME: warn if this is the case -- see comment about skipped	1
# TODO: this can probably be optimized	1
// TODO: Typecheck the parameters	1
// TODO: Kimish	1
// TODO: Order of this list is important as it affects type promotion. it's not	1
// Grab consumer domain entries and reverse replay map. TODO: Maybe	1
# TODO: do we want to test this too?	1
# 'fuse_fx', 'quantize_fx',  # TODO: add quantize_dynamic_fx	1
# TODO: maybe need more complex attr name here	1
// TODO: Remove this once the following issue is addressed:	1
# TODO: deduplicate annotation matching with Return	1
# FIXME: there are a few things that fall under this like	1
// TODO: rename context.h -> context_cpu.h & context_base.h -> context.h	1
# TODO: We might not need this anymore, since most scalars now show up	1
// TODO: Improve this by checking if it is mutated in the graph region	1
# TODO: make GivenTensor generic	1
# TODO: Deprecate and remove the following alias `_ConvTransposeMixin`.	1
// TODO: should serialize parameters with Module instead of with each Method.	1
// TODO: This constructor should probably use an ATen abstract method in order	1
# TODO: add support for more ops	1
// TODO: This is morally the same thing as KernelRegistrationConfig, but it's	1
# TODO: allow convert_custom_config_dict to override backend_config_dict	1
// TODO: Figure out how to unify these call interfaces.	1
// TODO: Consider a way to pre-allocate and reuse	1
// TODO: more vectorization with loop interleaving	1
// TODO After we actually export CALL instructions we can remove this.	1
// TODO: figure out why this needs to be computed...	1
// TODO: [algo] improve this algorithm, as it is horrendously inefficient.	1
# TODO: consider more complex/custom dynamic ranges for	1
# TODO: maybe update the cpp argument API to take optional namespace argument?	1
# TODO: miopen_LIBRARIES should return fullpath to the library file,	1
# TODO: should be `arg.type.is_tensor_like()`?	1
// TODO: vector may be faster	1
# TODO: If we ever implement tensor.nextafter, below is what we want ideally.	1
// TODO: is it ok that we're doing it eagerly? In the other implementation we	1
// TODO Make this explicit	1
# TODO: im not good enough with regexes to ignore -> *	1
// TODO: it's also odd these ops use gpu_kernel_with_scalars	1
// TODO: can we support caching this?	2
/// a graph (probably a dataflow graph).  TODO: refactor this	1
// TODO xinyu: standardrize reset_parameters virtual funcs	2
// TODO: should be input	1
# TODO: Remove this once script supports type() calls	1
# TODO: I think this is not necessary anymore	1
// TODO: Add constructors for all of the descriptors	1
// TODO: piping down the parallel dimension map here would	1
# TODO use arg_dequant_infos	2
// TODO: strides would also be important when we handle permutations in	1
# TODO: refactor to solve this weird dependency where	1
// TODO: Skip this if not writing tensors	1
// TODO: remove constant prop in the pass	1
# TODO: Use TypeAlias when Python 3.6 is deprecated	1
// TODO: _fused_dropout_cuda is to be removed, see PR #63937	1
# TODO (maybe): merge with embedding quantize handler	1
//! TODO: May want to sort this based on size of connections between this and	1
// TODO:  Run more tests for bsize > 128.	1
/* TODO: Use [[maybe-unused]] when C++17 becomes the standard */             \	1
// TODO: Support skipping python frames	1
* TODO: add docs after this is finalized.	1
// TODO: Consider a way to pre-allocate and reuse	1
"""FIXME: Temporarily replace std:: invocations of math functions	1
# FIXME: modernize these to be consistent with make_tensor	1
# TODO: include the configuration in backend_config_dict	1
# TODO allow kwargs such as unsafe and others for parametrization	1
// TODO: remove this when the cuda kernel is updated to support the channels_last memory format.	1
// TODO: this keeps reallocating map_size at every iteration, but we know	1
// TODO: Don't go through WrapRuntimeKernelFunctor	3
// TODO: we should have normal_like operator	1
# TODO: rename Relu -> ReLU to be more consistent with other classes	1
// TODO: There are places in core where a scalar is wrapped but not marked as	1
// TODO: make all operations that resize given outputs use this function	1
# TODO: should have an unspported list of operators, be optimistic for now	1
// TODO: a better invariant is that if we tagged, we MUST have a valid	1
// TODO: Change to fieldDesc	1
# TODO: change the way we get binary file -- binary may not in build/bin ?	1
ret = -1  # TODO: remove once JIT exceptions support control flow	1
// TODO: matrix-vector products in the code below are dispatched to matrix-matrix products.	1
// TODO: merge HasRand with CudaAnalysis.	1
# TODO: binary search	1
// TODO: Use the list from AMP eager directly	1
# FIXME: move to test_sparse or sparse utils	1
// TODO: Move to a dedicated validation pass	1
// TODO: support negative strides	1
// TODO: assert it's actually immediate	1
// TODO extend to support 4-bit qtensor.	1
// TODO: Consider representing debug info as a struct instead so you	1
# TODO: This can be further optimized by passing dim_in, dim_out = features,	1
// TODO: convert to schema, add a test	2
// TODO: verify whether source or dest device should be a priority in picking	1
//     TODO: Remove this constraint.	1
// TODO: Compare vs OpenSSL and/or CryptoPP implementations	1
* FIXME: use std::random_device with entropy information	1
// TODO: Change this when ChannelsLast3d is ready.	1
// TODO: Remove the condition on AT_ROCM_ENABLED entirely,	2
# TODO: TensorBase should work	1
// TODO: here is an optimization opportunity since welford uses int64_t for	1
# TODO: LSTM can't be TorchScript'd	1
// TODO: This test seems a bit goofy	2
// TODO: Dimension collapsing should be abstracted out and integrated into	1
//! TODO: Remove this interface as we do not intend to support dynamic	1
# TODO: do this with in-memory files as soon as torch.save will support it	1
# TODO: Fancier types?	1
# TODO: The unpacking is not yet implemented	1
# [old codegen] TODO: remove this? doesn't rename in codegen, it's just	2
// TODO: This error message seems awfully opaque	1
// TODO replace with TensorIterator implementation once #33166 is fixed.	1
// details. TODO Update to actually call pre-pack here once bias is removed	2
// TODO (viswanath): Should angle info be included as well while filtering?	2
//! TODO: In the next refactor PR, should put segment candidate	1
# TODO: use actual dtype instead of defaulting to float	1
// TODO: Fix this filter. Requires_grad is not the appropriate	1
# TODO: make this more efficient	1
# TODO: Re-enable this check (.type isn't supported in TorchScript)	1
# TODO: figure out what this does	1
# TODO: Implement `SeedSequence` like object for `torch.random`	1
/* TODO: result here is truncated to scalar_t,	1
// TODO: fast sigmoid	1
# TODO: move into common_utils.py or the test suite(s) that use this	2
// TODO: implement scale_grad_by_freq	1
# TODO: remove this (prefer make_symmetric_matrices below)	1
// TODO: remove this once we no longer have old TorchBind code	1
# TODO: avoid this special handling?	1
// TODO: try making the CUcontext thread local to see if that improves performance - why is this slow?	1
// TODO: enable if once shared libraries are unified in CMake	1
default = 'at::kLong'  # TODO: this is wrong	1
// TODO [unpickler refactor] __main__ isn't used by the pickler anymore, this	1
// TODO: remove this once the Kernel IR split is complete	1
// TODO: merge this code with case 1.	1
// manually. TODO: make declarable in native_functions	1
# TODO: include files like this should not set the default dtype	1
// TODO: these functions are unconditionally available because kaiser window depends on them	1
// TODO: merge this check up	1
// TODO: assert on empty buffer	1
// TODO: This line doesn't seem to be exercised at all in tests	1
# TODO: CUDA path doesn't work with batched or empty inputs	1
# TODO: reduce signatures down to one when optional args is available	1
// TODO: check the bounds only once	1
// TODO: [Need verify] looks like we can quantize simple functionals that just	1
inputs[4])})); // TODO: handle other dtypes of alpha and beta	1
// TODO: when we do legacy group convolution support, we'll repeatedly	1
// TODO add FBGEMM kernel	1
at::AutoDispatchBelowADInplaceOrView guard;  // TODO: remove	1
# TODO: Don't allocate a in-memory string for the protobuf	1
# TODO: directly translate a.default to python default	1
// FIXME: crappy implementation	2
# TODO: implement state_dict	1
// TODO: remove this custom tracing code once the custom op bugfix	2
# TODO: compute correct memory usage and CPU time once	1
// FIXME: assert ((D % VEC) == 0)	1
# TODO: skipping storage copy is wrong for meta, as meta	1
# TODO: test_cpu_gpu_parity doesn't handle case where output is not a singleton, submit fix	1
// TODO Make quantize_tensor_arm work for other datatypes too (int8, int32).	1
// TODO: are the `else` branches needed?	2
// TODO We don't call .cpu() on quantized tensors as it fails when calling	1
// TODO: This is marginally less efficient than it could	1
// TODO: generate a proper error log, as this probably means something	1
//! TODO:	1
// FIXME: remove magic > 0 after we ensure no models were serialized with -1 defaults.	1
// TODO: can be cleaner.	1
// TODO (T90180710): Simplify type_resolver and obj_loader when getting	1
// TODO: assert the output is a buffer and not a scalar	1
// TODO refactor so this function is usable both from jit and from aten	1
// TODO: For now, we are not moving the loads with the IfThenElse.	1
// TODO: handle the mask	1
//! TODO: cleaner way to set options?	1
// TODO: migrate all to using torchscript	1
// TODO Do schema inference without relying on WrapFunctionIntoRuntimeFunctor	1
// TODO: Remove duplicate declaration.	1
{})  # backend_config_dict, TODO: point to README doc when it's ready	1
ss << "{TODO implement Node::shape} ";	1
// TODO: maybe we should just push false and fallback	1
# TODO: Undo this special-case; see the header for motivation behind this	1
// TODO: Replace this helper with DECLARE/DEFINE_DISPATCH	1
# TODO: Add Pareto-Laplace KL Divergence	1
// TODO: call nvrtc.	1
// TODO @wconstab refactor using Symbol instead of string compare	1
// TODO: alias should be made aware to segmentation, so we'll always include	1
// TODO: Improved heuristic on when to coalesce or remove need to coalesce	1
// TODO: this is kind of... blegh	1
// TODO: when we fail commandline flag parsing, shall we continue, or	1
# TODO type process_group once `distributed` module is stubbed	1
// TODO: The assert is not necessary when we can handle matmul, right now we	1
// TODO: reduce the apparent redundancy of all the code below.	2
# TODO: Remove PYTORCH_MIOPEN_SUGGEST_NHWC once ROCm officially supports NHWC in MIOpen	1
# TODO: remove the skip after these two operators schemas are fixed	1
# TODO: add an API to map real -> complex dtypes	1
// TODO: Shall we handle the case when shape has -1 here?	2
// TODO Delete this once kernels don't do that anymore	1
# TODO: Remove the configuration by reference ('module')	1
// TODO: use LCM of stride and dilation to avoid unnecessary loops	2
# TODO: replace torch.divide with masked divide when available.	1
# FIXME: nansum does not support passing None to dim	1
# TODO: do we need eagerly calculate and save it here? Can it be derived	1
// TODO: remove, all constant tensors should have typed sizes	1
# TODO: consider removing this check and allowing users to specify	1
// TODO: These two functions below are slow!  Fix internal data structures so	1
# TODO: where	1
# FIXME: torchscript: div(float, float)	2
// TODO: Improve this once D31357486 is landed.	1
// TODO This is an incomplete implementation of std::apply, not working for	1
# TODO: return self	1
// FIXME: Allow any integral type.	2
// TODO if constexpr instead of enable_if	2
// TODO - relax this? CAFFE_ENFORCE_EQ(op->output_size(), 1);	1
# TODO define a bijection for LowerCholeskyTransform	1
# TODO: expand this to `_ConvNd` when channels_last support is extended	1
# TODO support future	1
// TODO we should attempt to call __str__ if the object defines it.	1
// TODO: investigate making this SingletonOrSharedTypePtr<TensorType>	1
# FIXME: second derivative is implemented but seems to be incorrect	1
// TODO: contiguous is called for further jit optimizations.	1
// FIXME: the shape of the input to the fictional PadPacked node has	1
# TODO: See https://github.com/pytorch/pytorch/issues/56285	1
// TODO: Make this take Variable by const reference	1
# FIXME: dim=None not supported	2
# TODO: Not sure why the arguments assigned here are for	1
# TODO: remove QConfigAny and replace it with Optional[QConfig]	1
// TODO:	1
# TODO remove and replace in favor of contextlib.nullcontext	1
# TODO: Missing type hints for nn	1
// TODO: setup grid-stride loop	2
# FIXME: amax reduces all dimensions when dim=[]	2
// TODO: check that weight matches output->sizes()	1
// TODO: Make a more type safe std::includes wrapper which disallows use	1
const float pred_ctr_x = ctr_x + width * dx; // TODO fuse madd	1
// TODO: This is a temporary apporoach to enable calling user fucntion	1
# TODO: maybe the logic to handle the legacy schema is no longer necessary?	1
# TODO: should use some canonical form instead of 'str(arg.type)' - see comments	1
# TODO @kiuk - make entrypoint a required field	1
// TODO This is an inefficient way to compite sign, and can be much faster	1
* TODO: t15868555 This algorithm is fast but can miss dependencies.	1
// TODO: move abs to aten namespace because it's schematized!	1
# FIXME: Remove when back testing is no longer required.	1
// TODO: guard impl_index, but I think that's not needed;	1
# TODO this should probably be a separate loss, not hacked in this one here	1
/* TODO: Use [[maybe-unused]] when C++17 becomes the standard */         \	1
// TODO: Adding this attr should be correct, but as of LLVM 9.0.1 adding it	1
// inplace argument is ignored now, TODO:support inplace	1
* TODO: This should be jettisoned in favor of `set_sizes_and_strides`,	2
// TODO: have to copy output because at::embedding doesnt have an out	1
// TODO: to really support input tensor large enought to go beyond int32,	2
// TODO: Turn on autocast by default. default turned off to avoid tests failures	1
# TODO: update notes/cuda.rst when this class handles 8+ GPUs well	1
# TODO is it correct to call_cur_module twice here?	1
// TODO Deprecate this function in favour of linalg_lu_factor_ex	1
// TODO: decide on fixpoint strategy	1
// TODO: avoid having to set this guard for custom mobile build with mobile	1
// TODO: implement prefetching if it starts mattering (TF does it)	1
# FIXME: prod does not support passing keepdim without passing dim	1
// TODO: remove this special case for HIP when issue is fixed:	2
# TODO: allow structured external backends later.	1
// TODO: restore decomposition after fusion, in case we are decomposing	1
# TODO: See https://github.com/pytorch/pytorch/issues/68592	1
// TODO: Expose in PyTorch Frontend	1
// TODO: It's a bit irritating that we have to do logical ORs here, it would	1
# TODO: consolidate this with the get_cases function from	1
// TODO: we could probably merge the two if it has perf impact on generated	1
# TODO: Once step_param interface is robust, refactor step to call	1
// TODO: Return const ptr eventually if possible	1
# TODO: only assertion error is bound in C++ compilation right now	2
/* TODO: remove once the bug is fixed. */                                \	1
// TODO: Remove once [serialization type tags] is landed	1
# TODO: Fix for python < 3.3	1
inline void WaitEvent(const Event& ev) { /* TODO */	1
# TODO: include the deprecation as soon as torch.testing.assert_close is stable	1
// TODO: this can be improved with summarizes of what the function does	1
// TODO: CPU instruction set selection should be folded into whatever	1
# TODO: add input, output validator	1
# TODO: Verify that sysconfig isn't inaccurate	1
// TODO: fix this when windows can correctly capture variables in nested lambda	2
TODO: test coverage	1
const auto output_padding = output_padding_arg; // TODO: Deconvolutions	1
// TODO find a way to parallelize this...	2
// TODO: use MaybeOwned	1
# TODO: check in THNN (if inplace == True, then assert value <= threshold)	1
# TODO: for _shared_model, no need to broadcast	1
// TODO: Check if the tensors with symbolic shapes are contiguous.	1
# TODO: complete the data type: bool, float16, byte, int64, string	1
# TODO not sure what this message really means	1
// TODO: we are having one unnecessary copy here if the context is already	2
TODO: To remove this check once Union support lands.	1
# TODO: FIXME	1
// TODO: add TORCH_API	2
// TODO: consolidate other ELF file related functions in loader.cpp to this file	1
// TODO: SetDeviceTensor accept vector<int64_t>	1
// TODO: figure out how to make compiler happy without dynamic casts	1
// TODO: replace filler distribution enum with a better abstraction	1
// TODO: support 16bit, 32bit, and etc.	2
thread=0)  # TODO: find in sqlite database	1
'ConstQuantizerPtr',  # TODO: rename	1
# TODO: how come ValuesView isn't a Sequence lol	1
// TODO: how to do this more intelligently	1
scalar_t scale = 1; // TODO: expose as argument?	1
# TODO: DivisiveNorm2d	1
// TODO: Currently tarjans mutates the graph, and that's the only reason we	1
// TODO: we don't support list type in codegen yet;	1
// TODO: exceptions in future	1
// TODO: Currently we do not support signal handling in non-Linux yet - below is	1
// TODO: if image decoding was unsuccessful, set label to 0	1
//! TODO: have a interface for grabbing all recent logs. Need to put a buffer	1
// TODO: move this from `at::` to `jit::torch::` after	1
# TODO: These invariants are weirdly asymmetric?	1
// TODO: More efficient would be to create event inside of main thread (at	1
# FIXME: `x` is a sparse view of `v`. Currently rebase_history for	1
// TODO: It might be good to use cpuinfo third-party dependency instead for	1
# TODO: Remove me once https://bugs.python.org/issue42666 is resolved	1
// TODO: only support NCHW for now	2
// TODO Change Ptr to DynamicTypePtr when all migrations are done.	1
// TODO: Expose this for real in ATen, some day?	1
// TODO: fix this when windows can correctly capture variables in nested lambda	1
# TODO: Consider not exporting these during wildcard import (reserve	1
// TODO: this might be slow - consider batched updates?	1
* TODO: need to support customizing equality	1
// TODO: Abstract stride logic to reuse with consumer indexing	1
// TODO: Add sorting options?	1
# FIXME: Undefined behavior sanitizer: shift exponent -9 is negative	1
TODO: when scale != 1 is introduced then use:	1
# TODO: It would be better to export this as a chunk directly, as this is	1
AutoDispatchBelowADInplaceOrView guard{}; // TODO: Remove.	1
// TODO: Re-audit this; it used to be an indexSelect directly into r_values	1
// TODO: c10::optional<>::value returns an rvalue ref so can't use it here??	1
# TODO update this when inplace namings are unified	1
# TODO: Consider adding a utility function to torch.jit to test	1
// TODO: switch to kUint64 when it is available.	1
false)); // TODO: nDim is bad, as it is collapsed	1
# TODO: maybe move to the generator side as it's not related to binding.	1
// TODO: gcc/clang has __builtin_clz() but it's not portable.	1
# TODO: currently model building does not have access to iter counter or	1
// TODO: remove	2
# TODO: enforce typing for each instance based on mode, otherwise	1
# FIXME: logical_xor does not accept scalar inputs	1
// TODO: dynamic allocation size: cur_rows*factor_j[i]*ranks[i+1]	1
// TODO: Remove this code in a separate diff, since we only have one	1
int stream_id = 0; // TODO: thread local stream id	1
# TODO: create weight observers from qconfig.weight	2
# TODO: Limitations and things about enable_python_mode we should fix before exposing it:	1
// TODO: modify this after resize_ added `memory_format` tag	1
# TODO: remove this (prefer make_symmetric_pd_matrices below)	1
// TODO: we need to stay on safer side instead of "default to return true	1
// TODO: extract arg_count from packed.	1
# FIXME (by @ssnl): Improve adaptive pooling docs: specify what the input and	1
# TODO: Validate supported requests	1
# TODO: Use a real parser here; this will get bamboozled	1
# TODO: ClassSimplexCriterion	1
# TODO: diagnostic if dir does not exist	3
// TODO: https://github.com/pytorch/pytorch/pull/59380#pullrequestreview-725310492	1
# TODO: some signatures of median do support out	1
// TODO: Remove once torchvision has been updated to use the ATen header	1
// TODO: No need to have this whole header, we can just put it all in	1
// TODO: compile the generated C++ kernel into a library,	2
// TODO Add tracing here	1
* FIXME: The behavior in this function is from legacy code	1
// TODO: support clamp_min, clamp_max	1
// TODO: test if Python key is disabled	1
// TODO use fast math when possible	2
// TODO: Change dims related arguments to int64_t?	1
// TODO: Build utility to strip off debug map. It should also do the	1
// TODO: When all kernels that use gpu_kernel_with_scalars are	2
// TODO: Simply traverse through uses from of. Would be a lot faster than	1
// TODO: implement prealloc optimization and fill in temp_sizes	1
* TODO: As people experiment with capture, keep an eye out for use cases that might need to	1
// TODO HIP support	1
/// TODO: it might be possible to handle cases where backward is	1
# TODO: Undo at least that second hack.  We should support string states.	1
# TODO: change the signature for fuser_method to take matched module patterns	1
// TODO: this message is not correct anymore, since this InferredType is	1
# TODO: switch to scale.item() after adding JIT support	1
OpInfo('trapz',  # TODO: in the future, 'trapz' should be made a proper alias of 'trapezoid'	1
// FIXME: Epsilon parameter isn't required, we don't take the reciprocal	2
// TODO Add GPU support by writing a generic wrapper.	1
// TODO: rand_like should support cast.	1
// TODO: deprecate?	1
// TODO this can be a qualified name check	1
# TODO: maybe we should change activation_post_process to _activation_post_process	1
// TODO add "canRunNatively" once memory management is audited	1
# TODO: Remove this scripting logic once the 2-week FC window has passed.	2
# TODO: If indexing is supported natively in ONNX in future opsets,	1
// TODO: we need to properly restore shape information after fusion.	2
# TODO: could add more detail here. For example, what the user should do	1
# TODO: Remove the try/except once all operators have sample_inputs_func with	2
// TODO: add fbgemm for per channel	1
# TODO: probably better to accumulate these errors and report them all	1
# TODO: specify __all__	1
// TODO: reuse memory for bufs with dynamic shapes	1
// TODO: check that output->size() matches output_sizes	1
// TODO: must be investigated and unified!!!	1
// TODO: [T87340633] Support reducing the batch dimension	1
# TODO: can we simplify this to always return a tuple of Tensor or None?	1
# FIXME: sum reduces all dimensions when dim=[]	7
// TODO: add in-place variant	1
// TODO: this function needs to be implemented and tested. Currently just throw	4
// TODO: if the divisor is a scalar, rewrite as multiplication by a constant.	1
// TODO: can't get rid of this use of TensorType	1
# FIXME: remove after implementing reflection pad 3d	1
// TODO Do schema inference without relying on WrapFunctionIntoRuntimeFunctor	4
// TODO: are the `else` branches needed?	2
# TODO: use slicing when slice optimization has landed	1
// TODO extend this special case to when the underlying storage of new_grad	1
// TODO: Handle tensors with different dtype, layout, device, memory	1
// TODO: numa?	1
// TODO: this should be a TypeError	3
// TODO: use op info to print out the op in a more user-friendly way	1
// TODO: investigate how "ExprPtr" can be implicitly converted to "ExprHandle"	1
// TODO: remove when script supports setting grad mode	5
// TODO: eliminate newCapacity.	1
// TODO: Deprecate these structs after we land this diff	1
// TODO - confirm that this is correct for NHWC	1
// TODO this needs to go in `m`s compilation unit	1
// TODO: How do we adjust this so we can reduce to a single scalar value?	1
# TODO: fail fast on quantization API usage error, then remove this class	1
// TODO: handle other alpha and beta dtypes, e.g. alpha=0.6, beta=0.2	1
// TODO: For future tasks, since output quantization parameters are set equal to	1
// TODO: make RRef system messages idempotent and retry on failures.	1
# TODO: dedup with BatchNorm2d	1
// TODO: This is a massive hack!  There is some confusion about	1
# TODO: add a copy kwarg that guarantees that the tensor is put into fresh	1
// TODO: Check what happens with MKL, the output error reported with non square matrices tends to be high	1
# TODO: maybe change to this when https://github.com/pytorch/pytorch/pull/32958 is landed	1
// TODO: We clone grad_slice because we modify it below and "fn" might save	1
// TODO: use CUDAGuard here instead of context and employ explicit sync	1
// TODO: channels last 3d	1
# TODO (zaf): Mask might not be part of the qconfig (T83295194)	1
// TODO include edges in the SCC in a smarter way.	1
// TODO: check latency here!!!!	1
# TODO Make return type more specific	1
# TODO: when standard argument type for "nets" is introduced,	1
// TODO This can probably use fused add multiply to get better perf	1
// TODO: is there a more idiomatic way to do this?	1
// TODO: Remove duplication with Upsample.h (CPU).	1
# TODO: shouldn't this be OptionalType[ListType[...]], since it defaults to None?	1
# FIXME: sum does not support passing keepdim without passing dim	1
#   TODO: some cpp naming logic (e.g. resolving name conflict) might be irrelevant?	1
* TODO: Should (2) and (3) be swapped?	1
// TODO: I think the answer is we shouldn't have used Symbol here	1
// TODO: need to see if there is extra error checking needed	1
# TODO: it would be nice to not have these special cases	1
// TODO: support runtime flag	1
# TODO: handle attrs	1
# TODO: we should pipe the exception of the failed subprocess here.	1
# TODO:	1
# TODO: Lambda for picking	1
// TODO: implement prefetching if it starts mattering (TF does it)	1
# TODO: instead of always doing this if there is an observer,	1
// TODO: Once the code in caffe2/python/onnx/backend.py no longer calls	1
TODO: fixme	1
// TODO possible to remove this arg by deferring the init value until we	1
// TODO: Tuning NumThreads for w_grad	2
// TODO: This function may be sub optimial. If we find that an iteration domain	1
// TODO: When we move to SM 3.5 we should update this	2
// TODO: This isn't right if there's a thread index at a higher level	1
// TODO Also register c10 operators on mobile	1
# TODO: Once we decide to break serialization FC, we can	4
// TODO: 5D channels last	1
// TODO: supports only single comprehension for now	2
# TODO: T18892922, use device annotations	1
# Special tree reduction for 16 gpus, TODO generalize like in muji.py	1
// TODO: Unify with DepTracker	1
# TODO: use argv	1
// TODO: handle dynamic dimension.	1
// risk of breaking existing clients. TODO: A better way would be to allow	1
# TODO: remove these once we support Type's in the JIT IR and we can once again	1
raise AssertionError("TODO not sure if there are other valid types to handle here")	2
// FIXME: make this thread-safe by reusing the benchmark cache in Conv_v7.cpp	1
# TODO: remove observed_op, looks like it's not used	1
// TODO: vectorization	1
# TODO: maybe don't represent default here	1
# TODO: This method has some duplicate lines with the	1
# TODO: maybe don't need keep scattered out fields for python signature?	1
# TODO: My kingdom for a pattern matcher	1
// TODO: in a follow up we need a global logging structure	1
// TODO: Update input names of function to match those in Module source code	1
// TODO: add filter to the clamp patterns and remove this pass	1
// TODO: Add remaining transforms	1
# TODO: compose all metas into one AssertionError	1
# [old codegen] TODO: because these aren't guaranteed to be 100% faithful	1
// TODO fix duplication caused by referencing same op across multiple	1
.device(at::kCPU) // TODO: support GPUs too	2
# TODO: compare structure (ensure analytic jacobian has correct shape)	1
// TODO: remove when serialization of dtype uninitialized tensor is removed	1
# TODO: type annotations for *args and **kwargs	1
// TODO: ThreadDim should be BlockDim and BlockDim should be GridDim	1
// TODO: consider to initializing to a blocked layout	1
# TODO: the second dim (num of input nodes) of param is after feature preproc,	2
// TODO: we frequently use pairwise root mapping from consumers to producers.	1
// TODO: Relax the checks to support dynamic shapes	2
# TODO: add a typing.Protocol to be able to tell Mypy that only objects with	1
# TODO: rename this to node_name_to_target_dtype_info	1
// TODO: Update this as well;	1
# TODO: Once we decide to break serialization FC, we can	3
// TODO: this should be really swapped for something more efficient	1
schema.overload_name().empty(), // @TODO: is this check correct?	1
// TODO: vectorize in accscalar_t?	2
// TODO: What if these parameters are not of the correct dimensionality?	1
//   TODO: remove d) from the requirements because the simplification formula	4
// TODO: We could allow freezing in this case but we would need to	1
// TODO: Land on a general solution for RPC ThreadLocalState. See	1
// TODO: add a macro to declare the filters	1
// TODO: consider to convert non-contiguous tensor to `ideep::tensor` directly.	1
# TODO check if we should set reduce_rage = True by default here	1
// FIXME: how do cases work?	1
// TODO: strides variant?	1
// TODO: Extend support to N-D batched embeddings, similar to	1
# TODO move op-specific logic out of here	2
# TODO use valid to mask invalid areas due to padding in loss	1
# TODO: this doesn't seem right...	1
// "aten::masked_fill.Tensor(Tensor self, Tensor mask, Tensor value) -> Tensor", TODO: requires 0-dim Tensor	1
int64_t model_version) { /* TODO: T90339189 deprecate all v3 when v3 models	1
// TODO: we don't care about merging multiple profiling runs as we don't	5
// TODO: Write gradient for this when needed	3
// FIXME: consider each updater above will broadcast its value with	1
// TODO: #ProfileIValue List should update this	1
# TODO: implement load_state_dict	1
# TODO: check that criterions don't ignore grad_output	1
// TODO: Stop manually allocating CUDA memory; allocate an ATen byte	1
// TODO: revisit complex inputs and equal_nan=true after	1
# FIXME: AssertionError: False is not true : Tensors failed to compare as equal!	2
* TODO: we could use different names for the following 'handle_torch_function'	1
// TODO: hacky way of determine the group size	1
# TODO refactor this code once we update the prepare logic to have additional information on	1
# FIXME: minimum does not accept scalar inputs	1
// TODO this faster solution does not work on Android build	1
// TODO: specialize for float2half2/half2float2?	1
// FIXME : code duplication with conv_dnnlowp_acc16_op.cc	1
// FIXME refactor aliasdb construction to be more robust to mutation so this	1
# (e.g., at::cpu::add).  We don't generate methods (TODO: do this	1
// TODO: Autotune/use better heuristics, improve speed more.	1
// TODO If just B requires grad, the following formula is better:	1
# TODO: backward uses in-place operations that vmap doesn't like	1
# TODO: Maybe this should be in tensor_classes? :)	1
// TODO: preserve the func type.	1
// TODO: rename to c10	1
// TODO: maybe consider deduplicating the definitions here, it's getting	1
# TODO: Possibly check scale and zero point.	1
# TODO: The above procedure does two matmul+allreduce steps per iteration --	2
# TODO: Conv2dLocal	1
// TODO: sameAs should have better logic to check against any type and return	1
// TODO: The implementation of `tensor_kernel_scan_outer_dim` and	1
// TODO: Replace me with inline constexpr variable when C++17 becomes available	1
# TODO move op-specific logic out of here	1
# TODO: implement timeout	1
// TODO: Remove once we clean up the GraphExecutor usage.	1
# TODO Consider using Q = torch.orgqr(*torch.geqrf(A)) to compute the Q of the QR _much_ faster	1
# TODO: core_trainer_sources is not necessary for libtorch lite	1
# TODO allow (loc,scale) parameterization to allow independent constraints.	1
// TODO Disallow this and rather use std::unordered_map/set everywhere	1
// TODO:	2
// TODO: memory stride should be considered here, our inference above is not	1
// TODO: Disable cont. branch to test more risky code	1
// TODO Make it work for more compilers	1
// TODO: handle optional out_qscale, out_qzero	1
// FIXME: don't do this if they're efficiently moveable.	1
# TODO add more checks	1
// TODO: Actually record which one we actually picked	1
// TODO: change the condition to `self_.dim() != 0` once we expose scalars	1
// TODO: expand this to convXd	1
// TODO: support the mask case	1
# TODO: Support context manager interface	1
// TODO: we're not using the most efficient algorithm here for simplicity.	1
// TODO: temporary hack to resolve my is_constructible issue;	1
// TODO: we could add __torch_function__ dispatch here but I don't know	1
// TODO: set correct domain for function proto.	1
# TODO: This is probably not exhaustive, but it's a start	1
# FIXME: uint8 input returns uint8 instead of bool	2
// TODO: Remove functions below when ChannelsLast3d is ready.	1
// TODO: the TensorIterator reduction implementation of mean	1
// TODO: The deprecation here triggers a deprecated use warning	1
// TODO: optimized kernel	1
# TODO: add try-except and destroy _agent in all processes if any fails.	1
// TODO: we probably have done this already up to this point	1
// TODO: change whether to include the parenthesis to the parent expression,	1
TODO: This is an inefficient implementation that uses `.dequantize`.	1
// TODO: Maybe check that compressed_size === file_size.	1
// TODO: Remove all these messages and use rpc + registered functions instead.	1
// TODO: Fix this pass/maybe get rid of this part.	1
// TODO Update quantize_tensor_arm implementation to follow quantize_val,	1
// TODO: replace self.inverse with linalg_inverse	1
// TODO: inference mode for chaining	1
# TODO: move this into library proper	2
# TODO: Add Exponential-Laplace KL Divergence	1
# TODO: explain this	1
// TODO: Memory use can probably be optimized by re-using kernels across GPUs with	1
// case we need to run aten ops (TODO: support different devices). The first	1
# TODO: Expand to remote RRefs.	1
// TODO: a possible optimization is to fuse the fp16 conversion into Concat	1
// TODO: restore the above, see https://github.com/pytorch/pytorch/issues/64709	1
# TODO currently placeholders/parameters aren't put into random partitions,	1
// TODO: trick the optimizer for case where C == 4?	1
// TODO: optimize in-place opererations and copy operations	1
// TODO: use adapter instead of istream?	2
# TODO: this does the wrong thing with KeyError	1
// TODO: I am not sure if we actually need the 'dropout' and 'train' parameters	1
// TODO: add GUARDED_BY once it is available	1
// TODO: remove this once we don't automatically enabled Autograd dispatch	1
# TODO: don't explicitly list dtypes here; get it from canonical	1
// TODO: this is an unnecessary copy. In theory we can directly	1
# TODO: handle fqn	1
* TODO: We may want to have ordering of outputs to inputs. I'm not sure why we	1
// TODO: Swtich to TensorIterator for better maintainablility and	1
// TODO What if it gets set later?	1
// TODO dim == 3 case will be enabled once it is fully tested	2
// TODO: call disarm rather than leak gil_scoped_acquired once PyThreadState_Clear can safely be called from finalize	1
// TODO: Make this structured to undo the perf regression from native:: removal	2
# TODO: FIXME: RuntimeError: "bitwise_or_cuda" not implemented for 'Half'	1
// TODO: remove after broadcasting is supported	3
# TODO: This feature could be added in the future	2
// TODO: remove this. This is a temporary list of functions that allow Python	1
// TODO: we could also parse extra arguments here and allow to pass in	1
1}); // TODO investigate how this is different from normal empty_strided	1
// TODO: better error message	1
// TODO: It's not good for these ops to be top-level, it makes cases	1
// TODO: enable with better TLS support on mobile	1
// TODO: keep running estimates.	1
# TODO: remove allow_list	1
// TODO: Actually, would this make ASAN's job harder catching	1
// TODO: mutex this guy;	1
CUDNN_RNN_ALGO_STANDARD, // TODO: verify correctness / efficiency.	1
# TODO: Pretty sure this approach loses ConstSequential status and such	1
# TODO: replace torch.maximum with masked maximum when available.	1
// TODO: Use lookahead to avoid creating the tuple and immediately	1
// TODO: this checks that the metavars occur directly as an index, but this	1
// TODO: remove when we have Type support in the IR	1
// TODO: could also take intersection of refinements present in	1
/* TODO: move this to a common place */	1
// TODO: Maybe show simple lists and tuples on one line.	2
// TODO: extend the supporting inputs here.	1
// TODO: code duplication with dnnlowp_op.h	1
// TODO holding this thing is creepy	1
// TODO: Try to avoid a copy here.	1
assert_jit_shape_analysis=False,  # TODO: support index.Tensor()	1
// TODO: Add more types (int32, int64)	1
// TODO: Need to redesign this part a bit to	1
// TODO: potential optimization - if there is a Symbolic	1
// TODO: This can be optimized.	1
# TODO: return self	1
# TODO: make DDP uneven inputs context manager support buffer	1
// TODO: These don't really belong here but torchvision builds in CI need them	1
// FIXME: code duplication with ConvDNNLowPOp::QuantizeBias_	1
// TODO: make this a context guard	1
# TODO: `cpp_type` is only to keep it byte-for-byte compatible with the old codegen, should remove.	1
// TODO handle multiple levels here	1
// TODO: In the ideal end state, it's okay to set disabled version_counter	1
// TODO: support shape inferencing. Right now we only handles static shape	1
// TODO: Swtich to TensorIterator for better maintainablility and	1
# TODO: see https://github.com/pytorch/pytorch/issues/64709	4
// TODO: Maybe show simple (empty?) dicts on one line.	1
// TODO: Check size, stride, offset, and numel and indicate if	1
// TODO: uncomment once unpack is implemented for BCSRMatrix	1
// TODO: proper implementation with masking.	2
// TODO: We have to extend it to support shapes vector.	1
// TODO: Consider generalizing this into a call stack.	1
// TODO: Proper fix is to create real descriptor classes	2
# TODO: Support non-equal-rank broadcast where semantics match.	1
# TODO MAKE SURE THAT DISABLING WORKS	1
// TODO:	1
// TODO: clean this up when https://github.com/pytorch/pytorch/issues/60306 is improved	1
// TODO: Complete after verifying utility of TT-layer's forward pass.	1
// TODO: consider interleaving herrmman merge and bruteforce merge, as	1
// TODO: this is in no way unique and is just a hack right now.	1
// TODO: add error reporting for graphs that can't be converted.	1
// TODO: descriptor checking	1
# TODO: move this to a separate function	1
// TODO: replace input.svd with linalg_svd when torch/xla can work with at::linalg_svd	1
// TODO: Make the Python API above to just call this C++ API.	1
// TODO: put this into the public API	1
# TODO: So far we don"t have a module using this method. We"ll keep	1
default_complex[0] = atof(str.c_str()); // TODO: parse "x + xj"?	1
// TODO: Remove this header	1
// TODO: Here can only use std::partial_sum for C++14,	1
// TODO: we should ideally be able to interrupt this blocking wait if we check	1
// TODO: remove the extra check when all the Tensors are properly initialized	1
# TODO: make helper functions for (torch.quint8, torch.qint8, None)	1
// TODO: trick the optimizer for case where C % 4 == 0?	1
# TODO: To remove this check once Union suppport in TorchScript lands.	1
// TODO: the above were the only checks in rnn.py, but it doesn't seem	1
# FIXME: numpy reference diverges: Comparing (nan+nanj) and (-0+0j)	1
# TODO (mingzhe09088): get rid of noqa	1
// TODO: will unify the two macros BUILD_LITE_INTERPRETER and C10_MOBILE soon.	1
# TODO: FIXME: sigmoid fails on complex inputs that require grad	1
// TODO: need to clean up all the env options	1
// TODO: Include alpha check for add/sub	1
// TODO: Support multiple quantization methods instead of assuming 2b1b.	1
# TODO: implement numpy-like issubdtype	1
// TODO: I think tensor geometry sufficient for weight_buf/weight	1
// FIXME: adding value comparison is slow	2
.DisallowInputFillers() // TODO: enable the filler	11
// TODO: Make this configurable.	1
# TODO: is this right? Don't really understand this	2
// TODO: move this to C10 and make it C10_API	1
// TODO: We should change parallelize interface to be on tensorview or at least	1
# TODO: reconcile with torch.linalg.det and torch.linalg.slogdet	1
// TODO: Ideally, this function would never be called if requires_grad is	1
// TODO: Deep-copy the module	1
// TODO: Can be changed to FIFO in order to avoid full traverse on every	1
# TODO: this is not handling non-tensor tuple args (for example,	1
// TODO: use llvm.abs intrinsic for LLVM 12	1
// TODO: when resize_cuda_ is re-written to be unified with resize_,	1
// TODO: move this to more generic location.	1
// TODO: extend to fusion of consumer into _producer's_ fusion blob	2
// TODO: remove context	1
// TODO: Refacto qnnpack_utils.h so as to separate code	1
// FIXME: This error is conservative. Detected an interface module	1
// TODO: resolve VarType if necessary	1
// TODO: This probably shouldn't actually be static inline	1
// TODO: once copy is exposed in Declarations.yaml we may be able to bind	1
// TODO: Refactor this so we just pass everything in via options	1
// TODO: Keep only the else branch once constant_folding is enabled by	1
// TODO: Should this be rfactor instead of root??	1
// TODO: This could be bad juju if someone calls globalContext() in the	2
// TODO: Test on GPUs.	1
# TODO: flatten allocates a std::vector, which could be expensive	1
// TODO this function makes broadcast communication call and	1
# FIXME: "prod_cpu" not implemented for 'BFloat16'	1
// TODO: some of these ops will not get generated because	1
// TODO: throw saved exceptions	1
// TODO: Can we change the logic of vectorizer so that we don't need this?	1
// TODO: currently only identify terms with one variable being mod; it is	1
// TODO: To achieve better performance we can have a pipe pool per	1
# TODO: put this somewhere else, maybe	1
# TODO: ConvTranspose2dMap	1
// TODO: this will return only the AccessInfo for A. It's included for	1
# TODO: for some reason weak_script_method causes a destruction of the	1
// TODO: this function is nontrivial and since CudaRTCFunction uses CRTP, it	1
// FIXME: pass in num_reduce_dims?!	2
// FIXME: this is a temporary solution to add a special-case for	1
# TODO: the following line needs to only check fqn	1
// TODO: what if the previous handler uses sa_sigaction?	1
// TODO: Use TensorGeometry here instead of the entire Tensor, which we	2
# TODO: remove the overriding implementations for LSTM and GRU when TorchScript	1
// TODO fields	1
// FIXME: There is no `operator<<` overload for `at::kBFloat16` type,	1
# TODO: I'm not sure if this counts as an implementation detail of	1
// TODO: Unranked SymbolicShape printing is ambiguous with that of	1
// TODO: the Python equivalent code has special-cased copy_to	1
# TODO: Improve this error message, possibly after converting	1
TODO:	1
// TODO: should we get symbolic_size instead and check for size	1
# TODO: SubtractiveNorm2d	1
# TODO: add different size support for sparse_nn_partition	1
// TODO: Verify that nodes in the pattern don't alias.	1
// TODO: Validate method_compile_spec.	1
TODO: The current implementation of this script only generates interfaces for built-in methods. To generate	1
// TODO: trying to reduce the variable number (common subexpression	2
// TODO (after Tensor merge) If we pass in a Blob holding a Tensor, extract	1
// TODO: use bitwise operator overloads once we add them	1
# TODO: create a new type of reducer with external weights to wrap	1
// TODO: too many things are currently abstracted under the term	1
// TODO: combine with TensorArg?  So far that's been for debugging, and this is functional...	1
# TODO: Fix test_out_arg_all_dtypes as torch.empty_like(expected_output) where expected_output=op(input)	1
# TODO: sm_backend_config_dict can fallback to use parent's backend_config_dict	1
# TODO: Add qat support for BNReLU2d	1
// TODO There is nothing in the system that relies on aten:: and prim::	2
# TODO mypy doesn't support @property, see: https://github.com/python/mypy/issues/6185	4
// TODO: make RRef an IValue, and edit createStackForSchema accordingly	1
// TODO: check scalarType	1
// TODO: reuse temporaries when possible (e.g. for inplace operations)	1
# TODO: this is hack to recognize NumberType	1
// TODO: Say what actually used it	1
// TODO: Some ops have conversion happen at Peephole pass.	1
// TODO enable fast handling for reductions	1
# TODO: Consider incorporating this into the data model	1
# TODO: handle errors here and just ignore the file?	1
# TODO: Either lint that GHA scripts only use 'set -eux' or make this more	1
# TODO: Add qat support for BNReLU3d	1
# TODO: make WEIGHT_INDEX_DICT and BIAS_INDEX_DICT an argument to the functions that needs them	1
# TODO: refactor & remove the following alias	1
_dim = [i for i in range(ndim)]  # noqa: C416 TODO: rewrite as list(range(m))	1
# TODO: FIXME: RuntimeError: "min_elementwise_cuda" not implemented for 'ComplexFloat'	2
// TODO combine this with quantize_val once the numerics for ARM are aligned	1
# TODO: Checking `ps.method and ('requires_grad' in parser_outputs)` is a hacky	1
// TODO torch::autograd::backward should take the c10::optional<Tensor> gradient directly	1
# TODO: investigate nondeterminism	1
with torch.onnx.select_model_mode_for_export(model, torch.onnx.TrainingMode.EVAL):  # TODO: move outside of torch.onnx?	1
// TODO (this will include the version number later)	1
// TODO: output(1) & output(2) should also be marked	4
// TODO: check type?	1
""" TODO (mingzhe): it is not necessary to sum up everything by myself,	1
// TODO: Now that set_output resizes both the original_tensor	1
// TODO: the following logic can be merged into regular Tensor class methods	1
// TODO: This function is not used.	1
// FIXME We should always extract DataPtrs, in order to catch the case of	1
// TODO: I think it may be possible to track inside the loop and	1
# TODO: Remove this scripting logic once the 2-week FC window has passed.	1
# TODO: no union or any types in TorchScript, make step a scalar tensor instead	1
main()  # TODO: Run this script automatically within the build and CI process	1
// TODO: - refactor and make explicit part of TE Kernel api	1
// TODO: Remove ROCM-specific behavior when https://github.com/pytorch/pytorch/issues/59750 is fixed.	1
* TODO: Deduplicate this with THTensor_(newWithTensor)	1
// TODO: put this into the library	1
// TODO: fast erf	1
# TODO: DDPSink is currently enabled for unused parameter detection and	1
// TODO: we can optimize dequantization by doing a premultiplication	1
// TODO: Use re2.	1
# TODO: Expose these directly to Python to avoid maintaining this list.	1
// FIXME: share version counters	1
# TODO: handle in place on tensor list	1
// TODO: is there a danger of us fusing operations that's supposed to be on	1
# TODO: make this part of something more general, or get rid of it.	1
# TODO: delete these special cases; see tools.codegen.api.cpp--these	1
// TODO: The algorithm below can probably be optimized.	1
// TODO: Replace me with std::numbers::pi when C++20 is there	1
// TODO: should this run through dispatch on this and other?	1
# TODO: These are the modules that cannot be observed	1
// TODO: contiguous is called for further JIT optimizations.	1
# TODO: verify that weight comply with TRT structured sparsity requirements:	1
// TODO: support for exceptions	1
// TODO: Not compatible with embedding dim larger than maxThread	1
// TODO: This at::abs() call is used so that the at::abs() call in the	1
// TODO: this is currently too wide. It detects whether a store-target	1
// TODO: try passing the "mapped" file directly to cuModuleLoadCall instead of using an intermediate buffer	1
// TODO: use PackAMatrix if filter_qparams_[0].zero_point == 0	1
// TODO: it would be nice if we could use	1
# TODO: move these out	1
# TODO: maybe orgnize this better (e.g. break down to more functions)	1
* FIXME: You can move this function to Generator.cpp if the algorithm	2
// FIXME: Consider assigning over existing elements, rather than clearing &	1
// TODO: we should comply to codegen type promotion.	1
// TODO: it's possible to make the _out variant to be a primal function and implement linalg_eigh on top of _out	1
// FIXME: stride should be optional	1
// TODO Making operator== noexcept if underlying type is noexcept equality	1
// TODO: replace (module, method_name) with graph?	1
// TODO: handle scaling factor when it's not constant 1;	1
node_->kind() == aten::cat, "TODO: generalize logic");	1
// TODO: multi-scale histogram for this thing	1
// TODO: is weight required to be contiguous?	2
// TODO: if necessary, use dispatcher.	1
// TODO Support more types	1
* TODO: we could consider exposing API to allow custom registration of parsing	1
// TODO: Why?! Can't we just flip the order here...	1
// TODO: figure out how to commonize this with int8 quantize	1
// TODO: provide guarantees that compiler won't optimize this out	1
# TODO: Now, there is something interesting going on here.  In the code below,	1
// TODO: add invariants	1
# TODO generalize this for more things	1
// TODO: we should probably get a list that's close to what our fuser handles	1
# TODO: RuntimeError: While computing batched gradients,	1
# TODO: support more than just LSTM	1
// TODO: out op variants	1
# TODO: add qat.Conv1d	1
//TODO handle multi-return functors	1
# TODO: eliminate mask_input as unnecessary when using masked divide.	1
// FIXME: this should be (and was) Scalar::toTensor, but there is currently no way	1
// TODO: to support more data type	1
// TODO: different iterator types for sizes & strides to prevent	1
// TODO: use current device id from thread local instead of passing gpu in	1
// FIXME: Current Alias analysis fails to track subvalues.	1
// TODO: uncomment the following when passing incorrectly sized 'result' is not allowed	1
#include <iostream> // TODO: remove, debugging only	1
// TODO: The kernels are copied from fbgemm_gpu, we should dedup them later	2
// TODO: replace with at::zeros when it's implemented for sparse csr	2
// TODO: Enable view in parser by detecting non-alias view operation	1
// TODO: insert CUDA's async stream waits; tracing and counters	1
// TODO: Add List[bool] once .to<c10::List<bool>> doesn't throw an error	1
// TODO: if we fail GlobalInit(), should we continue?	1
// TODO: check that listeners are not relying on prepareForDeregistration()	1
// TODO this isn't a scalable way to determine parallelism.	1
# TODO: Check tensor types for ops	1
# TODO: make into assert	1
# TODO: Remove string escape once Python-3.6 no longer supported	1
// TODO: output 3 & 4 are not created	1
// TODO: check for casting once it's supported	1
# TODO: Add Beta-Laplace KL Divergence	1
// TODO: we should make the tradeoff here to use thread_local instead of global	1
# TODO: this list might be incomplete.	1
// TODO: Move this to fixed_divisor.h	1
# TODO: may need to change the key to Node regenerate the map in each transformation,	3
# TODO: Use a real parser here; this will get bamboozled	1
# TODO: remove special case for operator.getitem	1
// TODO: once modules are first class in the interpreter and methods are not	1
// TODO: we can consider preallocating and pre-filling the args vector.	1
// TODO: Remove template?	1
// TODO remove after TensorOptions rationalization	1
// TODO: better error message	1
// TODO: handle cases where we need to generate > 2^32 element tensors	1
# TODO: Move this into ONNX main library	2
# TODO: Add ContinuousBernoulli-Laplace KL Divergence	1
// TODO: refactor this file (one per namespace)	1
// TODO: Remove once we fully migrate to non-blocking mode.	1
# TODO: See if we can extract GPU vs CPU information from the PyTorch model	1
// TODO: named constructor to avoid default initialization	1
TODO: @kefeilu: this function's body should be moved into the actual calling	1
# TODO use glob	1
// TODO: this can return MaybeOwned	1
# FIXME: bfloat16 backward support likely depends on CUDA11+	2
// TODO: Replace the link to the documentation once it's available.	1
# TODO: delete this list once we make all nn_tests work	1
# TODO: This could get us in recomputation trouble if b.expr is nontrivial	1
// TODO: We should add the following, but we need to go through schedulers	1
// TODO: fix fb internal use-case so that it doesn't trigger this internal assert when the base is not a view.	2
// TODO this is deprecated but we don't throw a warning because a lot of ops in	1
// TODO: (@krshrimali) Try inheriting from TensorIteratorBase instead.	1
# hardcoded for now, TODO: expose the api to user,	1
// TODO: Set to 1?	1
// TODO: Tuning NumThreads for sum_squares	2
// TODO: remove?	1
Callable,  # fp32 op type (TODO future PR: add quantized op type)	1
// TODO: masked_scale_cuda is to be removed, see PR #63937	1
//! TODO: this can be largely expanded to look at complete	1
// TODO: Clean up what remains here	1
# TODO: Manually add `self.param_groups` if using a functional	1
# TODO: proper overriding analysis when implementing class inheritance	1
// TODO: Extend support for attribute of type List[Tensor] etc.	1
// TODO: generalize logic to for other tensor input ops when they are	1
// TODO: Currently we only support (*, Sparse) combination for (tensor.layout(), tensor.grad.layout())	1
// TODO: should we support runtime compilation with updated dynamic shape;	1
// TODO: even though this API is currently used **only** in codegen to	1
// TODO: consider storing namespace separately too	1
// TODO: allow abstract kernels to reuse generated kernels from common pool	1
#                       TODO: test coverage for mixed types inputs.	1
// TODO: When Variable is added, delete these constructors	1
/// TODO: This function encourages bad behavior (assuming CUDA is	1
// of graph inputs. TODO: remove	1
# TODO: FIXME: complex inputs requiring grad error in forward	3
* TODO: Look into changing the threading semantics of Generators in ATen (e.g.,	1
// TODO: eliminate this conditional when zero-size dims supported correctly	1
// TODO: find a better metric in using ldg or not. Support different dtypes.	1
// FIXME: remove const_cast once unary_op_impl_out is updated	1
# FIXME: improve precision	6
// TODO: avoid spilling W by breaking out the non-padded vs padded case.	2
# TODO: we don't currently do this functions that are recursively	1
// TODO: Blegh, bare references	1
// TODO: could transform up to 2 other dims in the same cuFFT operation	1
// TODO: Use liveness analysis to catch more general scenario	1
# TODO: byte-for-byte compatible with old codegen behavior - it's incorrect to assume	1
// TODO (Ashkan): Disabling temporarily.	1
// TODO: Assert it's an ATen identifier???	1
// TODO: assert all provided preferred roots are in the history of reference	1
# TODO: FIXME: cholesky_inverse throws an error in forward when requires_grad=True	1
# TODO: merge this with elementwise bench	1
# TODO: @krshrimali, once to_numpy method in SampleInput class is modified to take None inputs,	1
# TODO (add more FileCheck signature)	1
// TODO: eliminate me	1
# TODO: fix LSTM handling in eager mode static quant and remove this	1
// TODO: what's an "inverted interval"? Open on the left	2
// TODO: Should this actually be in launch params?	1
# TODO: dedup this branch	1
// TODO: See if it's possible to use those directly.	2
// TODO Use fbgemm kernel to pack values	2
// TODO: This is a temporary approach to allow C++ thread to correctly	1
# TODO: clean up old codegen behavior	1
# TODO: support regex as well	1
// TODO: Could consider putting some of	1
# TODO: move the exceptions to proper locations	1
// FIXME: Not actually doing floor division (#43874)	2
# TODO: Might need a fix in torch group_norm module	1
// TODO: for now don't attempt partial factorization of this	2
# TODO fix when https://github.com/python/mypy/issues/2427 is address	1
# TODO: look into using weakref here instead.	2
# TODO (zaf): Inherit from `quantized.LinearPackedParams` (T83294430)	1
// TODO: no ifs for now	2
# TODO: this WAR is for https://github.com/pytorch/pytorch/issues/18524	1
# TODO: move DefaultFuseHandler	1
# TODO: Maybe make these names match the original.	1
# TODO: some signatures of nanmedian do support out	1
// TODO: try to remove this	1
// TODO: more user friendly API	3
# TODO: rename this file to config_utils	1
# TODO: Remove this once ScriptModule supports registering None buffer	2
// TODO: we can avoid this guard by moving operations	1
# TODO: do not run this twice on input and output	1
// TODO: get rid of temp memory use	1
// TODO: Maybe add better logging here.	1
// TODO: support clamp_min.Tensor(Tensor self, Tensor min) -> Tensor	1
// TODO: this step would not be deterministic, because valuesBetween isn't	1
// TODO: Remove once we migrate everything to non-blocking mode.	1
// TODO: DebugUtil will be upstreamed after LazyTensor is in.	1
// TODO: Ideally we only add AutogradBackend key when the tensor requires	1
int64_t model_version); /* TODO: T90339189 deprecate all v3 when v3 models	1
// TODO: `aten::sum` is too flexible, we should restrict for a better	1
# TODO: add limited pickling support for sharing an iterator	1
# TODO: fix bug in the documentation for svd_lowrank:	1
"TODO: support more cuDNN activation modes");	1
// TODO: get std::forward<> to work	1
// FIXME This does some unpickling, which could be a bit expensive:	1
// TODO: support color jitter and color lighting in gpu_transform	2
# TODO: switch to zero_point.item() after adding JIT support	1
// TODO We still save ws is because of the current design of workspace and	1
// TODO: let's check the data type for buffer and skip if it's good	1
# TODO: make attentions a generic state	1
// TODO: For real input, perform rfftn then mirror with conjugate symmetry	2
// TODO: check if objects have been freed from time to time	1
# TODO: may need to change the mapping when we support dynamic quantization	1
// TODO: allow abstract kernels to use multiple generated kernels	1
* found matches, no nodes in the subgraph alias with each other). TODO: check	1
// TODO: is_reduction is too hacky here. we should categorize operation types	1
// TODO this leads to ambiguous cases (NC11) to be always treated as contiguous	1
// TODO: fill?	1
// TODO: Maybe check that only the batch dimension is changed?	1
// TODO: handle epilogue	2
// TODO: be more explicit about the full key set at call sites so we	1
// TODO remove the mutation here	1
// TODO: rename flags to C10	1
# TODO: maybe make "pattern" to be a list of patterns	1
# TODO: look into rewriting with early return and getting loop unrolling to fire	1
///  // TODO: when tensors are stored in the pickle, delete this	1
# TODO: check gradients for parameters, not just inputs	1
# TODO: When PyTorch drops the support for Python 3.6, it can be converted	1
// TODO : init value support or remove.	1
// TODO: Look into using DepthFirstGraphNodeIterator	1
// TODO: Deprecate this instancecheck entirely.  It's here to make	1
// FIXME: warn if this is the case	1
# TODO: MSECriterion weight	1
# TODO More precise types here.	1
# TODO: if statement only here to tell the jit to skip emitting this when it is None	1
# TODO make these types more precise	1
// TODO This whole file should be deleted and replaced with the mechanism	1
constexpr int thread_work_size = 4; // TODO: make template substitution once we decide where those vars live	1
// TODO: std::vector<int> -> std::vector<int64_t>	1
# TODO: this doesn't work with RNN ops	1
# TODO: This is a little inaccurate, because it will also pick	1
// TODO: consider storing shape compute graph directly on operator,	1
// TODO: if we extend TensorIterator to accept 3 inputs,	1
# TODO: Should handle optional here?	2
// TODO: need to raise an error when you impl a function that has a	2
// TODO: Handle the storage_order properly to get the NCWH.	1
// TODO: change it to CUDAGuard	1
// TODO: Would be great if we didn't need this.	1
// TODO: TO be removed, once this properly works from libkineto	1
# TODO: clamp shares tensors among its sample inputs --- we should prohibit this!	1
// TODO: why is this optional?	2
// TODO: for_blob produces non-resizable tensors, we might want this to be	1
// TODO: make this more const correct	1
// TODO: The new CUDNN_BATCHNORM_SPATIAL_PERSISTENT mode was	1
// TODO: Need to bump 'num_expected_refs' here when we support post_hooks for	1
# FIXME: clone indices on construction.	1
// TODO: Maybe add better logging here.	29
* FIXME: Remove support of the legacy state in the future?	1
// TODO: only use this if necessary (add a pass to find all shared ivalues,	1
// TODO: The Python version also accepts arguments	1
// TODO: Check stride and indicate if the tensor is channels-last or non-contiguous	1
// TODO: Implement precison-based formatting	1
// TODO: Seriously consider writing the derivative formulas for	1
// TODO: change to a proper error report	4
# TODO: make use of reduce like below when JIT is ready with the missing features:	1
# FIXME: ldexp does not accept scalar inputs	1
// TODO: merge this loop below.	1
// TODO: [T87567124] Fully implement chunk in Metal shader	1
# TODO: ContrastiveNorm2d	1
// TODO: add a check for OpArgCount and op_type	1
* TODO: We may want a BFS version of this code to extract ILP, not implemented	1
// TODO make this an iterator and simply promote it on insertion.	1
// TODO: Remove annotateInputShapes pass when TraceGraph can also capture	1
// TODO: remove this, there are no codegenerated checks for devices yet	1
// Zero-fill undefined grads (TODO: do this more efficiently)	1
// TODO: the operator depends on output being set to 0 before the run	1
// TODO: is it worth optimizing this loop via padding in C?	1
//  TODO: merge with the binary counter parts	1
// TODO: No version; c.f. https://github.com/Maratyszcza/NNPACK/issues/165	1
// TODO: if neither *at nor *bt is 1, ensure they are identical	1
// TODO: either decompose composite ops like slice or add handling here	1
return at::dequantize(out); // TODO: optimized kernel that outputs fp32 so	2
// TODO: Review const model, and objects	1
// TODO: debugging mode to see the qualifier.  We definitely	1
raise AssertionError("TODO not sure if there are other valid types to handle here")	1
// TODO: Revisit this if Bufs need to be cloned as well.	1
// TODO add factor as an input, need to check Split::Split during validation	1
// TODO: This is a temporary solution. We should pass enough information to	1
// TODO: check perf vs dedicated kernel.	2
// TODO: raise warning when parsing deprecated signatures	1
// TODO: make mkldnn tensor serialize...	1
// TODO: There is no need to branch with every element	1
// TODO: getDevice() ?	1
// TODO: make a version that takes an impl argument.  Unfortunately,	1
// TODO: parallel_method->getParallelType	1
// TODO: do XLA	1
// FIXME Isn't it too verbose for a library to print logs in normal operation?	1
# TODO Remove once gloo submodule is recent enough to contain upstream fix.	2
# FIXME Once we consolidate the error messages returned by the	1
SparseHIP, // TODO: I think this is not actually used, due to Note	1
// TODO (tugsuu) move this calculation into a seperate step.	1
// TODO: we are following the convention for no good reason;	1
# TODO:	2
# TODO: torch.complex32 when properly supported	1
// TODO: calculate the version_token.	1
# TODO: currently we hard code the root node, which only works for	1
// TODO: Shouldn't all returned results be successful?	1
# FIXME: count_nonzero does not accept keepdim kwarg	1
# TODO: I guess we should do copyreg too?	1
# TODO: Because torch.jit._IgnoreContextManager relies on Python's `exec` method	1
# TODO: for ops with structured_delegate it should check the dispatch table of	1
// TODO: consider TLS (tid + tls counter)	1
// TODO: review if this is computing in double when given a float input	1
// TODO: In C++17 we should be able to use the filesystem header.	1
# TODO: This is to keep same byte-for-byte result as the old codegen - maybe unnecessary?	1
// TODO: Decide what kind of fixed point strategy we will have	1
# TODO: FIXME: lcm doesn't support scalars	1
// TODO: Think of a smarter way to leverage tbb::thread_arena to limit the	1
//TODO maybe think about unifying offset calculators and reuse	1
// TODO: simplify	1
// TODO: see if this pass can be replaced with peephole pass	1
# TODO: Need to benchmark the performance of lowering linear as fully_connected versus	1
// TODO: dont require dimensions of tensors to be set AOT ?	1
// TODO: merge this check up	1
// TODO - smallvector here ?	1
// TODO: This is awful code.  Also it doesn't work on Windows.	1
# FIXME: the docs say that persistent_id should only return a string	2
// TODO: update this to become a static assertion	1
# FIXME: bfloat16 backward support likely depends on CUDA11+ and SM53+	1
# TODO: refactor this to use iterate_and_apply	1
// TODO: Make traversal items local to this function.	1
// TODO: Inline intermediate operations (avoid inlining unrolled/vectorized	1
TODO: maybe rename this to TensorValueOpQuantizeHandler	1
# FIXME: both derivatives are implemented incorrectly	1
// ReductionSchedulerMultiDimNonFastest TODO: test reenablement to make	1
// TODO: should be inputs	1
// TODO: Abstract stride logic to reuse with producer indexing	1
# TODO: state dict offloading, activation offloading	1
// TODO: measure which default value will give better	1
// TODO: Improve this once D31357486 is landed.	1
// TODO (zach): we should consider skipping tensor factories in the cases	1
// TODO: this is not needed as we are only using the first val	1
// TODO: Make this an aten function and replace as_strided_qtensorimpl once that is done.	1
// TODO: support cast of output types	2
# TODO: Talk to ONNX about unconditional cast of scalar to float	1
# FIXME: Undefined behavior sanitizer	1
# FIXME matmul(x,y) + bias currently goes through jit AD, and backward formula in AD is not optimized for this	1
# TODO: byte-for-byte compatible with old codegen behavior - should clean up	1
# TODO: Merge these two templates together in the future once TorchScript syntax is improved.	1
// TODO: unique_name	2
// TODO: Make this more sophisticated. A value being the same as another value	1
# TODO: Figure out if this is safe. It seems like when generating the type signatures for	1
# TODO: remove this as onnx opset 11 spec allows negative axes	1
# TODO: fix this special case in PythonArgParser?	1
# TODO: Annotate with TypedDict when 3.8 is the minimum supported verson.	1
input.ndimension() == 4 &&   // TODO: 5-D contiguous depthwise is not supported yet, need benchmarks	1
# TODO after stopping workers, wait at least monitor_interval*2 for	1
//!  TODO: This implementation looks at common producers only, since common	1
// TODO: another copy paste from jit, refactor so it's usable from both	1
# TODO: rename to something more general	1
# TODO: Stop hardcoding that the output type is a Tensor.  Note	1
# TODO: evaluate optimizing this if needed.	1
// TODO: Remove this global var	1
// TODO: Do these make sense?	1
# TODO (refactor) this is duplicated, maybe have a helper function	1
# TODO: support check for standalone module	1
self.model = model  # TODO: Need to figure out how to load without this.	2
// TODO: would be nice if there were easy facility to look at uses and see	1
// TODO: check the op_type and make a real decision	3
// TODO: Instead of going up the loop nest we should go through the indices in	1
// TODO: Enable view in parser by detecting non-alias view operation	1
// TODO: What if inputSizes is not of the expected dimensionality?	1
# TODO: namespace threshold in 'nn'	1
// TODO: Return an ArrayRef instead (and delete the singleton while you're at	1
// TODO: Our current linear mode impls use unbound indices	1
// TODO A previous implementation of alias analysis always accessed	1
// TODO: out_axis_dim is assumed to be the same as the extent of	1
// TODO: To support all features of MemoryFormat::Preserve we need to add	1
// FIXME: use occupancy calculator instead	1
# TODO: handle non-tensor inputs	2
# FIXME: would be nicer if TensorOptions was optional based; not adding default arguments for options given	1
// TODO: I'm not really sure if we're actually obligated to traverse PyObject	1
// FIXME: if input < kVecSizeInFloat, can't vectorize at all	1
// TODO: fetch scalar type from Tensor? But it doesn't really matter...	1
# TODO: Add support for comparing meta tensors. See https://github.com/pytorch/pytorch/pull/67032.	1
// TODO: Consider putting the stub definitions in another class, so that one	1
// TODO: contiguous can be made to preserve the memory format	1
// TODO: extract & guard profile_ivalue; but how do we restore it???	1
// TODO: include all the other ways of adding these args.	1
// TODO: we will need a kernel_ir cloner to make this	1
# TODO: improve error propagation	1
# TODO: something smarter	1
at::AutoDispatchBelowAutograd guard;  // TODO: remove	4
// TODO: fbgemm::Quantize doesn't support taking in the	1
// TODO: Are we sure these tensors will always come into this fucntion with the	1
# TODO: merge with default static mapping	1
# TODO: Consider defining some aliases for our Union[...] types, to make	1
for (const auto i : c10::irange(N)) { // TODO: multithreading	1
TODO: add a README when it's more stable	1
# TODO: Provide more useful diagnostics.	1
// TODO Reuse stack vector instead of allocating?	1
# FIXME: sum does not support passing None to dim	1
# TODO: consider time-bound constraints as well.	1
// TODO: Improve the performance of this by figuring out a better approach.	1
// TODO: preallocate `runArgs` during compilation and fill in values where	1
// TODO: RpcCommandBase should have an abstract execute() method that we can	1
# TODO: test_non_contiguous_tensors doesn't handle case where output is not a singleton (such as	1
# TODO: Remove this when `make_tensor` supports excluding `0`.	1
# TODO: FIXME: jiterator does not support casting to complex outs	1
// Half. TODO: __truncdfhf2	1
# TODO: This only works with new style gcc and clang (not the old -faddress-sanitizer).	1
// TODO to support single process multiple devices and multi device modules,	1
// TODO: we might need to do two pass to avoid adverse memory allocations	1
# FIXME: isclose does not accept scalar inputs	1
// TODO: include this as warning once we have a more consolidated	1