https://github.com/pytorch/pytorch summary

[jaykchen]

fullgraph=True doesn't actually raise error when you don't manage full graph inside DDP https://github.com/pytorch/pytorch/issues/107639 The central problem discussed in the issue is that when the first graph fails to compile, an error is not raised and the code continues execution. The main solutions proposed or agreed upon are not mentioned in the post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

[DDP PT2] TypeError: convert_frame_assert.._convert_frame_assert() missing 2 required positional arguments: 'hooks' and 'frame_state' https://github.com/pytorch/pytorch/issues/107637 The central problem discussed in the issue is a TypeError that occurs when calling the function convert_frame_assert with missing positional arguments 'hooks' and 'frame_state'. The user tried to locate the missing arguments but couldn't find their origin. The issue is related to a model using DistributedDataParallel.

The main solutions proposed or suggested in the issue are not explicitly mentioned in the given information. However, based on the context, potential solutions could include checking the function call hierarchy to ensure all required arguments are provided, investigating if any recent changes caused the issue, or reaching out to the mentioned individuals for assistance.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

nvfuser does not respect CMAKE_INSTALL_PREFIX when build (cmake) libtorch https://github.com/pytorch/pytorch/issues/107633 The central problem discussed in the issue is that the nvfuser library does not respect the CMAKE_INSTALL_PREFIX when building libtorch using CMake. This results in the "libnvfuser_codegen.so" file being missing from the installation directory. The suggested solution is to make the nvfuser library's install destination settings consistent with other libraries in order to ensure that the file is installed correctly.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

torch.fx.Interpreter modules don't get compiled https://github.com/pytorch/pytorch/issues/107631 The central problem discussed in the issue is that the torch.fx.Interpreter modules are not getting compiled due to being skipfile'd. The user prefers using torch.fx.Interpreter over the torch.fx Python codegen for better stack traces and ease of debugging. One solution proposed is to generate the Python codegen instead, which would allow graph break functionality. However, the user believes that at least one of their reasons for using torch.fx.Interpreter (stack traces) could be made to work with Python codegen. The issue of graph break functionality is not considered a big deal as it is mostly used for models without graph breaks.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

torch._dynamo.exc.InternalTorchDynamoError: 'NoneType' object has no attribute 'guards' https://github.com/pytorch/pytorch/issues/107630 The central problem discussed in the issue is a bug related to the torch._dynamo.exc.InternalTorchDynamoError, specifically the error message "'NoneType' object has no attribute 'guards'". The user provides steps to reproduce the bug, along with a backtrace and a link to the full dynamo debug log. The user also mentions the versions being used and tags relevant individuals. No solutions or agreements are mentioned in the post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

ModuleNotFoundError: No module named 'torchgen.code_template' https://github.com/pytorch/pytorch/issues/107627 The central problem discussed in the issue is a 'ModuleNotFoundError' for the module 'torchgen.code_template'. The user encountered this error while doing Android benchmarking using a PyTorch tutorial. The main solutions proposed or agreed upon are not mentioned in the given information.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Incorrect code snippet for SavePlanner https://github.com/pytorch/pytorch/issues/107619 Issue: Incorrect code snippet for SavePlanner

The central problem discussed in the issue is that the provided code snippet does not return the expected plan. It seems to have an issue with its implementation.

The main solutions proposed or agreed upon are not mentioned in the provided content. The post only mentions the GitHub handles of several users who might be involved or interested in the issue.

The issue requires further analysis to understand the specific problem and potential solutions.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Previous version not found https://github.com/pytorch/pytorch/issues/107611 The central problem discussed in the issue is that the user is unable to download a specific version of PyTorch (1.8.0 with CUDA 11.1) using the provided command. The error message suggests that there is no matching distribution for the requested version. No solutions or alternatives have been proposed or agreed upon in the issue.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Support AMD Smart Access Memory https://github.com/pytorch/pytorch/issues/107605 The central problem discussed in the issue is that the user's Asus Zenbook S13 OLED, which runs on AMD Ryzen 6800U APU with 680M Graphics Card, is facing Out of Memory (OOM) situations when using Stable Diffusion PyTorch in Linux. The shared memory of the graphics card is set to a default of 512MB, and the BIOS setting does not allow users to modify the dedicated memory. The user is requesting support for AMD Smart Access Memory in PyTorch to dynamically change the memory size like in Windows.

One potential solution mentioned in the issue is the use of the "torch-apu

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Fx Graph nodes generated via aot_module have empty 'stack_trace' attributes https://github.com/pytorch/pytorch/issues/107599 The central problem discussed in the GitHub issue is that when using aot_module to generate Fx Graphs, the node.stack_trace attribute is empty, which prevents the user from referencing the original source and understanding the context of the constants used.

The main solutions proposed are:

1. Wrapping my_compiler with aot_autograd and passing it to torch.compile using the backend="aot_eager" option.

2. Using aot_function or aot_module_simplified instead of aot_module, as they are intended to have better support for user-facing

                   this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

torch._dynamo.exc.Unsupported: call_method UserDefinedObjectVariable(defaultdict) items [] {} https://github.com/pytorch/pytorch/issues/107595 The central problem discussed in the issue is a bug occurring in the torchrec FusedEmbeddingBagCollection, specifically in the torch._dynamo.exc.Unsupported code. The proximal cause of the bug is identified as the keys of the defaultdict being tuples of enums. No main solutions or agreements are mentioned in the available information.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

dynamo: don't graph break on ctx.mark_dirty https://github.com/pytorch/pytorch/issues/107593 The central problem discussed in the issue is the lack of support for torch.autograd.Function::ctx.mark_dirty in Dynamo, which currently causes a graph break. The user is requesting to add Dynamo support for this function without causing a graph break. The issue is related to the training UX for Float8.

No specific solutions or agreements are mentioned in the issue post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

repeat_interleave does not support tensor indexes on different devices while repeat does https://github.com/pytorch/pytorch/issues/107591 Issue Title: repeat_interleave does not support tensor indexes on different devices while repeat does

Central Problem: The issue is related to the ONNX export of a model using repeat_interleave with dynamic shapes. The Tensor.repeat_interleave operation requires tensors to be on the same device, unlike Tensor.repeat. This causes a RuntimeError when the indexes of the repeats are on different devices.

Solutions Proposed: One potential solution is to modify the ONNX export process to ensure that the my_tensor.shape[0] is treated as a Python-native int instead of a Tensor on the CPU

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Select on a coalesced COO tensor returns COO tensor with coalesce flag set to False. https://github.com/pytorch/pytorch/issues/107590 The central problem in this issue is that when performing a select operation on a coalesced COO tensor in PyTorch, the resulting tensor is not always coalesced, even though it should be. The ordering of indices in the select result should be the same as in the input tensor indices.

The main solution proposed is to ensure that the result of the select operation on a coalesced COO tensor is always coalesced. The expected result should be True in this case.

No specific agreement or further solutions were mentioned in the issue. The issue was labeled as "module: sparse" and

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Run transformers.OPTForCausalLM(config=config) occurs 'GraphModule' object has no attribute 'compile_subgraph_reason' https://github.com/pytorch/pytorch/issues/107587 The central problem discussed in the issue is that when trying to convert the whole opt model fw/bw without any break, the error 'GraphModule' object has no attribute 'compile_subgraph_reason' occurs. This issue is similar to a previously closed issue in the pytorch repository, but it has not been resolved in the nightly version. The post does not mention any proposed or agreed-upon solutions.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

[FakeTensor] to doesn't error with allow_non_fake_inputs=False https://github.com/pytorch/pytorch/issues/107582 The central problem discussed in the issue is that the to function in the FakeTensor module does not produce an error when allow_non_fake_inputs is set to False. The main solutions proposed or agreed upon are not mentioned in the given information.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

[LibTorch/iOS] Building with METAL support script is freezing https://github.com/pytorch/pytorch/issues/107581 The central problem discussed in the issue is that when trying to build the LibTorch library with METAL support on iOS, the build script freezes at 86% or gets stuck at 99% completion.

The main solution proposed by the user 'hadiidbouk' is to ensure that the 'build_ios' directory is empty before running the build command. They suggest editing the script to remove the contents of the 'build_ios' directory. Additionally, the user suggests using a different build command: 'BUILD_PYTORCH_MOBILE=1 IOS_ARCH=arm64 ./scripts/build_ios.sh'. However, it is not

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Doc is unclear on how to install pytorch with Cuda via pip https://github.com/pytorch/pytorch/issues/107580 The central problem discussed in the issue is the lack of clarity in the PyTorch documentation on how to install PyTorch with CUDA using pip. The user expressed frustration with the documentation not providing clear instructions for this installation method. One potential solution suggested is to delete the relevant installation parts in the documentation and link to the interactive installer on the pytorch.org website. Another suggestion is to add a link to the interactive installer in the documentation. The user also mentioned the need for clearer language in the documentation. Overall, the issue revolves around the need for clearer instructions and improved documentation for installing PyTorch with CUDA via pip.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

halo，I continue pretrain llama2-13B model ，but save state_dict is about 50GB file https://github.com/pytorch/pytorch/issues/107575 Issue: The issue reported by user 'moonmon' is related to the size of the saved state_dict file when continuing the pretraining of the llama2-13B model. The user mentions that the saved file is about 50GB in size.

Potential Solutions: There are no specific solutions proposed or agreed upon in the issue. However, one user suggests double-checking what is actually being saved to clarify the situation. The user also mentions that a 50GB file size is not surprising considering the size of the llama2-13B model and the use of the Adam optimizer.

Overall, the issue revolves around the large file size of the saved state_dict and potentially finding ways to optimize or reduce the file size.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

caching keys+values in TransformerDecoderLayer for faster inference https://github.com/pytorch/pytorch/issues/107573 The central problem discussed in the issue is that the current nn.TransformerDecoderLayer (and Encoder) does not support caching of keys and values, which results in inefficient inference during autoregressive generation. The inference currently takes O(k^2 d) time complexity for each step, but with caching, it can be reduced to O(k d).

The main solution proposed is to add an option to feed in previously calculated keys and values, so that unnecessary computations can be avoided. The issue raises the question of whether this feature should be included in the core or be provided as an example for users to follow. The comment from msarouf

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

DISABLED test_restore_range_dynamic_shapes (main.DynamicShapesSubGraphTests) https://github.com/pytorch/pytorch/issues/107572 The central problem discussed in the issue is that the test 'test_restore_range_dynamic_shapes' is failing in the PyTorch CI. The issue has been labeled as 'triaged' and 'skipped', indicating that it has been acknowledged and temporarily disabled. The issue provides debugging instructions to study the logs and mentions recent instances of failures and successes. The main solution proposed is to disable the test for the platforms 'linux' and 'slow' in the PyTorch CI.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

RuntimeError: Unsupported value kind: Tensor while torch.jit.script nn.Module https://github.com/pytorch/pytorch/issues/107568 Issue: 'RuntimeError: Unsupported value kind: Tensor' while torch.jit.script nn.Module. User is trying to script Lovasz loss function but encounters an error that TorchScript does not support Tensor. They provide code and environment information.

Potential Solutions: No specific solutions are mentioned in the post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Running DDP with IPV6 address, the client can not connect to the address. https://github.com/pytorch/pytorch/issues/107563 The central problem discussed in the issue is that when running DDP processes with an IPv6 address, the client is unable to connect to the address. The main solution proposed or agreed upon is to check the network configuration and ensure that the IPv6 address can be connected in the elastic launch environment.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Dynamo guards on unused Tensor variables https://github.com/pytorch/pytorch/issues/107561 The central problem discussed in the issue is that Dynamo is generating unnecessary guards on unused Tensor variables. These guards are not needed and can be pruned from the compiled graph. The main solutions proposed or agreed upon are not mentioned in the given information.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Floating point exception when running torch.nn.AdaptiveMaxPool3d https://github.com/pytorch/pytorch/issues/107557 The central problem discussed in the issue is a floating point exception error when running the torch.nn.AdaptiveMaxPool3d function due to arg_1_1 and arg_1_2 being set to None. The issue is reported with the specific versions of PyTorch, CUDA, and other relevant libraries. The main solutions or agreements proposed are not mentioned in the provided information.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer multiplication overflow when running torch.nn.AdaptiveAvgPool2d https://github.com/pytorch/pytorch/issues/107556 The central problem discussed in the issue is an integer multiplication overflow when running the torch.nn.AdaptiveAvgPool2d function. The main solution proposed is to use a different data type, such as float, to avoid the overflow.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer multiplication overflow when running torch.nn.MaxUnpool3d https://github.com/pytorch/pytorch/issues/107555 The central problem discussed in the issue is an integer multiplication overflow when running torch.nn.MaxUnpool3d with a specific value for arg_1. The main solutions proposed or agreed upon are not mentioned in the provided information.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer multiplication overflow when running torch.diagflat https://github.com/pytorch/pytorch/issues/107554 The central problem discussed in the issue is an integer multiplication overflow when running the torch.diagflat function. The user shared a specific value (arg_2 = 1250999896764) that triggers this bug. However, there is no clear solution proposed or agreed upon in the issue.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Storage size calculation overflowed when torch.nn.Upsample https://github.com/pytorch/pytorch/issues/107553 The central problem in the GitHub issue is that there is an overflow error in the storage size calculation when using the torch.nn.Upsample function. The user provided the argument arg_1 = 1250999896765, which seems to trigger this issue. The main solutions proposed or agreed upon are not mentioned in the provided information.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Storage size calculation overflowed when running torch.nn.functional.interpolate https://github.com/pytorch/pytorch/issues/107552 Issue Summary:

The central problem discussed in the issue is that the storage size calculation overflows when running torch.nn.functional.interpolate due to a very large integer variable. The user provides information about the PyTorch version, CUDA version, and other relevant libraries.

No solutions or agreements are mentioned in the post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Floating point exception when running torch.nn.functional.adaptive_max_pool3d https://github.com/pytorch/pytorch/issues/107551 The central problem discussed in the issue is a "Floating point exception" error when running torch.nn.functional.adaptive_max_pool3d. The main solution suggested is that the issue might be due to a negative input tensor.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer multiplication overflow when running torch.eye https://github.com/pytorch/pytorch/issues/107550 Issue: 'Integer multiplication overflow when running torch.eye'

Problem: The bug is caused by a very large integer argument when running torch.eye in PyTorch version 1.12.0. This leads to an integer multiplication overflow.

Potential Solutions: No specific solutions or agreements have been mentioned in the issue post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer calculation overflow when running torch.nn.functional.adaptive_avg_pool2d https://github.com/pytorch/pytorch/issues/107548 Issue: 'Integer calculation overflow when running torch.nn.functional.adaptive_avg_pool2d' Problem: The bug occurs due to large integer list elements. Solutions: The issue does not provide any specific proposed solutions or agreements. The user only provides code and log messages.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer overflow when running torch.nn.functional.upsample_bilinear https://github.com/pytorch/pytorch/issues/107546 The central problem discussed in the issue is an integer overflow error when running the torch.nn.functional.upsample_bilinear function. The issue is caused by a very large integer variable. No specific solutions are mentioned in the post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer overflow when running torch.nn.functional.upsample https://github.com/pytorch/pytorch/issues/107545 The central problem in the GitHub issue is an integer overflow occurring when running the torch.nn.functional.upsample function due to a large integer value. The user provides information about the bug and their system configuration. No specific solutions or agreements are mentioned in the post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer overflow when running torch.nn.ReplicationPad3d https://github.com/pytorch/pytorch/issues/107544 The central problem discussed in the issue is an integer overflow occurring when running the torch.nn.ReplicationPad3d function. The user 'dmc1778' encountered this problem due to a very large integer variable. There is no specific mention of proposed or agreed-upon solutions in the issue. Therefore, the potential solutions remain unknown.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer overflow when running torch.nn.AdaptiveAvgPool2d https://github.com/pytorch/pytorch/issues/107543 User 'dmc1778' opened an issue titled 'Integer overflow when running torch.nn.AdaptiveAvgPool2d'. The bug is caused by a large integer value. No further details about the issue, such as the reproducible code or the log message, are provided. No solutions or agreements have been proposed or discussed in the issue.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Inconsistent results when running torch.lu_solve https://github.com/pytorch/pytorch/issues/107542 The central problem in this GitHub issue is that there are inconsistent results when running the torch.lu_solve function on both CPU and GPU. The user suspects that this inconsistency is caused by a large negative input tensor. No specific solutions or agreements have been proposed or discussed in the issue.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer overflow when running torch.nn.MaxUnpool2d https://github.com/pytorch/pytorch/issues/107541 The central problem discussed in the issue is an integer overflow that occurs when running the torch.nn.MaxUnpool2d function due to feeding large negative integer values. The main solution proposed or agreed upon is to modify the code to handle negative integer values properly.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Index out of bound when running torch.gather https://github.com/pytorch/pytorch/issues/107540 The central problem discussed in the issue is an index out of bound error when running the torch.gather function in PyTorch. The issue is likely caused by a very large negative input tensor. No specific solutions are mentioned in the post, but potential fixes could include handling the negative input tensor differently or adjusting the code to avoid the index out of bound error.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Integer overflow when running torch.nn.functional.max_unpool2d https://github.com/pytorch/pytorch/issues/107539 The central problem discussed in the issue is an integer overflow error that occurs when running the torch.nn.functional.max_unpool2d function. This error occurs when feeding large input tensors or negative integer values.

One potential solution proposed by user 'msaroufim' is to consolidate all the similar issues into one. This suggestion is made to facilitate a more focused and effective discussion on the problem.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

inconsistent results with fx graph inplace updating cached constant https://github.com/pytorch/pytorch/issues/107538 The central problem discussed in the issue is that when using the fx tracing feature in PyTorch, inplace updates to a tensor that is cached as a constant can lead to inconsistent results. The user suggests two potential solutions: either recognizing when a tensor constant is inplace updated in the graph and using torch.clone in that case, or adding an explanation of this behavior to the documentation.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Ran into flatbuffer verifier problem while compiling tag v2.0.1 in win64 https://github.com/pytorch/pytorch/issues/107537 The central problem discussed in the issue is a flatbuffer verifier problem encountered while compiling tag v2.0.1 in win64. The user mentions that compiling version 2.1.0 worked, suggesting that it may not be a fundamental error. The main solution proposed is to ensure that the runtime libraries used match the version of code generated. The user asks whether these runtime libraries should be included in the repository or obtained elsewhere. Another suggestion is to check if the flatbuffers headers in the specified directory are compatible with the usage site mentioned in the error log. The user also suggests refreshing or updating the third-party submodule references in PyTorch.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Appending new logs to existing tbevent files when using tensorboard https://github.com/pytorch/pytorch/issues/107536 The central problem discussed in the issue is the lack of capability to append new logs to an existing tfevent file when using tensorboard. This makes it difficult to maintain a consistent log of progress and creates multiple files in the directory.

The main solution proposed is to add the capability to append new logs to an existing tfevent file using the SummaryWriter. This would allow users to continue logging to a single event file across multiple runs, ensuring a seamless record of training progress.

No alternatives or additional context were provided in the issue.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

NNPACK slow down M1/M2 Mac CPU https://github.com/pytorch/pytorch/issues/107534 The central problem discussed in the issue is that on ARM Mac (specifically M2), torch versions 1.12.0 and higher with NNPACK enabled are significantly slower compared to torch versions 1.11.0 and below. Disabling NNPACK in torch 1.12.0 results in similar performance as torch 1.11.0. The proposed solutions are to either disable NNPACK by default on these ARM architectures or develop better backends for M1/M2 CPUs.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

ReferenceError: weakly-referenced object no longer exists https://github.com/pytorch/pytorch/issues/107531 The central problem discussed in the issue is a 'ReferenceError: weakly-referenced object no longer exists' that is causing TorchAudio's CI to fail. The issue occurs on Python versions 3.8, 3.9, and 3.10. The main solutions proposed or agreed upon are not mentioned in the post.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Inconsistent results when running torch.arctanh https://github.com/pytorch/pytorch/issues/107528 The issue titled 'Inconsistent results when running torch.arctanh' discusses discrepancies in results between CPU and GPU. A user reported inconsistent outputs on different devices. One comment mentioned that they couldn't reproduce the issue with the latest nightly version. The central problem is the inconsistent behavior of torch.arctanh on different devices. The suggested solution is to update to the latest nightly version.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Export of quantized::linear_relu operator not supported with torch.onnx.export https://github.com/pytorch/pytorch/issues/107522 Issue: User 'finlytics-hub' opened an issue titled 'Export of quantized::linear_relu operator not supported with torch.onnx.export'. The user is unable to export a custom quantized torch model with fused modules.

Potential Solutions: No specific solutions are mentioned in the post. The post mainly provides information about the bug, including stacktrace and system configurations.

                    this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

conv2d wrong results on 3090/3090ti https://github.com/pytorch/pytorch/issues/107514 Issue: 'conv2d wrong results on 3090/3090ti'

Crux: When using torch.backends.cudnn.allow_tf32 = True on NVIDIA 3090/3090ti GPUs, the results of the conv2d operation are incorrect and have a larger error compared to float32. Setting torch.backends.cudnn.allow_tf32 = False resolves the issue. The problem is due to the rounding behavior of tensor cores used in tensor operations.

Solutions:

1. Set torch.backends.cudnn.allow_tf32 = False to disable tensor cores and preserve precision during convolution

                   this result is generated by flows.network. Triggered by @jaykchen

[jaykchen]

Increase Compatibility with Keras Core https://github.com/pytorch/pytorch/issues/107511 Issue Title: Increase Compatibility with Keras Core User: AlexanderLavelle Central problem: The issue discusses the need to increase compatibility between PyTorch and Keras Core to enable PyTorch native users to easily use their existing code with Keras Core. Main solutions proposed: The main solution proposed is to address the incompatibility issues with losses and embeddings between PyTorch and Keras Core. To increase the speed of running by utilizing a "jax" backend, the PyTorch ops that are not accounted for in JAX need to be replaced by more generic Keras Core ops. Summary: The issue highlights the

                    this result is generated by flows.network. Triggered by @jaykchen