// Note: we hold onto mu_ for the rest of this function. In practice, this
    // is okay, as stat requests are typically fast, and concurrent requests are
    // often for the same file. Future work can split this up into one lock per
    // key if this proves to be a significant performance bottleneck.
    absl::MutexLock lock(&mu_);
    if (LookupLocked(key, value)) {
      return TF_SetStatus(status, TF_OK, "");
    }

//    "/job:localhost/replica:0/task:0/device:GPU:1"}
// Then executing an operation on CUSTOM:0 will execute it on GPU:0 and GPU:1.
//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//
// All component tensors must have the same dtype. Currently they must also have

  // once unless the tape is persistent) and produces the gradient of the target
  // tensors with respect to the source tensors. The output gradients are used
  // if not empty and not null. The result is populated with one tensor per
  // target element.
  Status ComputeGradient(
      AbstractContext* ctx, absl::Span<AbstractTensorHandle* const> targets,
      absl::Span<AbstractTensorHandle* const> sources,