"@local_xla//xla:status_macros",
        "@local_xla//xla/client:local_client",
        "@local_xla//xla/pjrt:pjrt_client",
        "@local_xla//xla/pjrt:pjrt_future",
        "@local_xla//xla/pjrt:pjrt_stream_executor_client",
        "@local_xla//xla/pjrt:tracked_device_buffer",
        "@local_xla//xla/service:shaped_buffer",
        "@local_xla//xla/stream_executor:device_memory_allocator",

        [],
        otherwise = [
            "@local_xla//xla/stream_executor/cuda:all_runtime",
            "@local_xla//xla/stream_executor/cuda:cuda_driver",
            "@local_xla//xla/stream_executor/cuda:cuda_platform",
            "@local_xla//xla/stream_executor/cuda:cudnn_plugin",
            "@local_xla//xla/stream_executor/cuda:cufft_plugin",
            "@local_xla//xla/stream_executor:cuda_platform",
        ],
    ),

#     --config=dbg --per_file_copt=+tensorflow/core/kernels/identity_op.*@-g
# Since this .bazelrc file is synced between the tensorflow/tensorflow repo and
# the openxla/xla repo, also include debug info for files under xla/.
build:dbg --per_file_copt=+.*,-tensorflow.*,-xla.*@-g0
build:dbg --per_file_copt=+tensorflow/core/kernels.*@-g0
# for now, disable arm_neon. see: https://github.com/tensorflow/tensorflow/issues/33360

      llvm::cl::desc("Choose target opset."),
      llvm::cl::values(
          clEnumValN(OpSet::TF, "TF",
                     "Uses TF ops that mimic quantization behavior"),
          clEnumValN(OpSet::XLA, "XLA", "Uses TF XLA ops"),
          clEnumValN(OpSet::UNIFORM_QUANTIZED, "UNIFORM_QUANTIZED",
                     "Uses TF Uniform Quantized ops"))};

  Option<bool> enable_per_channel_quantization_{

}

def TF_TPUPartitionedOutputOp : TF_Op<"TPUPartitionedOutput", [Pure]> {
  let summary = [{
An op that demultiplexes a tensor to be sharded by XLA to a list of partitioned
  }];

  let description = [{
outputs outside the XLA computation.
  }];

  let arguments = (ins
    TF_Tensor:$inputs,

    DefaultValuedOptionalAttr<I64Attr, "0">:$partition_dim,

        "@llvm-project//mlir:Support",
        "@llvm-project//mlir:TransformUtils",
        "@llvm-project//mlir:Transforms",
        "@local_xla//xla:status",
        "@local_xla//xla:statusor",
        "@local_xla//xla/mlir_hlo",
        "@stablehlo//:stablehlo_ops",
    ],
)

cc_library(
    name = "tensorflow_lite_optimize",
    srcs = [
        "transforms/generated_optimize.inc",
        "transforms/optimize.cc",

                ' of compute capabilities excluding version %s.' % ver)
            all_valid = False
          if ver < 35:
            print('WARNING: XLA does not support CUDA compute capabilities '
                  'lower than sm_35. Disable XLA when running on older GPUs.')
      else:
        ver = float(m.group(0))
        if ver < 3.0:

// have a valid XLA type.
LogicalResult CheckClusterResults(mlir::tf_device::ClusterOp cluster) {
  for (OpResult result : cluster.getResults()) {
    if (!tensorflow::TypeValidForXLA(result.getType())) {
      return cluster.emitError()
             << "The ExtractHeadTailOutsideCompilation pass produced a Device "
                "cluster with a result with a non-XLA type: "
             << result.getType();

  }
};

// This pass performs a manual conversion with FakeQuant, converting between
// floating point and quantized space. It is designed to reproduce TF's
// implementation, mirroring the previous XLA implementation.
//
// 1. Computing proper quantized bounds. This involves nudging the input bounds.
// 2. Converting the input bounds to quantized space, rounding values.
// 3. Convert back into floating point space.

#include "tensorflow/compiler/mlir/tensorflow/utils/verification_utils.h"
#include "tensorflow/compiler/mlir/tf2xla/transforms/legalize_tf_with_tf2xla_passes.h"
#include "tensorflow/compiler/mlir/tf2xla/transforms/passes.h"
#include "xla/mlir_hlo/mhlo/IR/hlo_ops.h"

#define DEBUG_TYPE "tf-tfl-legalization"

namespace mlir {
namespace TFL {
namespace {
// Returns a TF_CastOp to I32. This function is used for CastOps that are