Tracker for TensorFlow-DirectML

[penpornk]

cc: @PatriceVignola @wchao1115

This issue tracks pending PRs, issues, and possible cherry-picks necessary for TensorFlow-DirectML for each TF release. Please post a comment with new things to track and I will update this post to reflect the changes.

New PRs:

• Fix duplicate kernel registration errors when using a Pluggable Device with the GPU type #56707
  • Awaiting internal discussion in the TensorFlow team.
• Return an error code for TF_GetInputTensorFromVariable #55677

PRs that need more investigation.

• [PluggableDevice] Add DEVICE_DEFAULT int32 registration for strided slice ops #55382 [A similar PR broke tests before so this can't go in until the issue is resolved.]
• [PluggableDevice] Add DEVICE_DEFAULT int32 registration for Pack #55381
  • Reverted because it broke internal tests.

PRs that made it into TF nightly (post r2.10 branch cut):

• [PluggableDevice] Add TF_IsHostMemoryInput and TF_IsHostMemoryOutput to the C API #51759
• [PluggableDevice] Make graph_buf argument in TF_NewFunctionLibraryDefinition constant #55544
• [PluggableDevice] Add TF_KernelBuilder_Label #55557
• Remove TF_CAPI_EXPORT from TF_InitGraph #55579
• [PluggableDevice] Add TF_OpKernelConstruction_GetNodeDef #52157

PRs that made it into TF 2.10:

• Add TF_AssignRefVariable #55379
  • Resubmitted in Add TF_AssignRefVariable #55640
• Add a validate_shape argument to TF_AssignVariable #55678
• [PluggableDevice] Add TF_AddNVariant and TF_ZerosLikeVariant #55645
• Add DEVICE_DEFAULT registration for DataFormatDimMap and DataFormatVecPermute #55558

PRs that made it into TF 2.9:

• [PluggableDevice] Add DEVICE_DEFAULT registration for ZerosLike and OnesLike #55173
• [PluggableDevice] Enable tensor list kernels on Windows #54468
• [PluggableDevice] Make candidate_input_indices constant in TF_ForwardInputOrAllocateOutput #54139
• [PluggableDevice] Add range checks to TF_ExpectedOutputDataType in non-debug builds #51758
• [PluggableDevice] Add DEVICE_DEFAULT int32 registration for Slice #55384
• [PluggableDevice] Add DEVICE_DEFAULT int32 registration for Concat and ConcatV2 #55385
• [PluggableDevice] Move ZerosLike DEVICE_DEFAULT registration outside of CUDA ifdef #55386
• [PluggableDevice] Move Unpack DEVICE_DEFAULT registration outside of CUDA ifdef #55387
• [PluggableDevice] Move Min DEVICE_DEFAULT registration outside of CUDA ifdef #55392
• [PluggableDevice] Move Max DEVICE_DEFAULT registration outside of CUDA ifdef #55393
• [PluggableDevice] Add int32 DEVICE_DEFAULT registration for Where #54746
• [PluggableDevice] Add TF_IsRefInput #55395

Closed PR:

• Add error handling to TF_GetInputTensorFromVariable #55362

[PatriceVignola]

@penpornk How can pluggable devices implement the Assign op which uses the legacy variables? TF_AssignVariable already exists to implement AssignVariableOp which uses resources, but there doesn't seem to be an equivalent for the legacy variables. Should we add a TF_AssignRefVariable function?

[penpornk]

@PatriceVignola

@penpornk How can pluggable devices implement the Assign op which uses the legacy variables? TF_AssignVariable already exists to implement AssignVariableOp which uses resources, but there doesn't seem to be an equivalent for the legacy variables. Should we add a TF_AssignRefVariable function?

Looping in @wangpengmit for more info on variables.

[wangpengmit]

Resource and ref vars are very different, so my first thought is that TF_AssignRefVariable is probably needed.

[PatriceVignola]

I have a working prototype in a fork that I put in the kernels_experimental header. I can submit a PR if that's ok with everyone.

Also, my understanding is that ref vars are deprecated in TF2 and are replaced by resources at the python API level. Is that correct? But even though they are deprecated, some popular benchmarks like AI-Benchmark use a frozen TF1 model when running on TF2, which yield bad results for pluggable devices since they don't currently support them.

[wangpengmit]

my understanding is that ref vars are deprecated in TF2 and are replaced by resources at the python API level. Is that correct? But even though they are deprecated, some popular benchmarks like AI-Benchmark use a frozen TF1 model when running on TF2.

Yes, ref vars are deprecated at the Python level. Existing TF1 models may still be using them, so TF2 internals still support them.

[penpornk]

I have a working prototype in a fork that I put in the kernels_experimental header. I can submit a PR if that's ok with everyone.

@PatriceVignola Oh, that's great! Please submit the PR and we can continue the discussion there (e.g., whether this needs to be an RFC, etc). Thank you very much! :)

[PatriceVignola]

@penpornk I submitted a PR here: #55379

[PatriceVignola]

New PRs to add to the list:

#55381
#55382
#55384
#55385
#55386
#55387

[PatriceVignola]

3 new PRs:

#55392
#55393
#55395

[PatriceVignola]

@penpornk A simple PR that was opened a month ago. I think it flew under the radar.

#54746

[PatriceVignola]

New PR: #55544