Hailo10H Compiler: Layer Support for Input Tensors with Large Dimension Sizes

While compiling EfficientLoFTR using Hailo AI SW Suite 2026-01, I have encountered the following error, when compiling the reshape operation:

[1, 6400, 12, 12] → [1, 6400, 1, 144] ≈ [1, C·F, H, W] → [1, C·F, 1, H·W] — fails

[error] eloftr Multi Context Partition Failed (duration: 0s)

[error] eloftr Partition and Allocation Failed (duration: 34s)

[error] eloftr Compilation Failed (duration: 34s)

No successful assignments: format_conversion68_defuse_width_feature_reshape errors:

Agent infeasible

failed on kernel validation: LCUReshapeWidthFeaturesTranspose does not support

    two subclusters with one apu in current arch

failed on kernel validation: format_conversion68_defuse_width_feature_reshape

    has 2 APUs but max allowed is 1

[error] Failed to produce compiled graph

As the reshape op I have implemented is supported by the DFC compiler according to the Guidelines (5.5.29). I wonder this error lies in the high size of the tensor in certain axis. The same error: “2 APUs but max allowed is 1”, showed up for a 2d conv with similarly sized tensor.

Thus my questions are:

  1. Does the problem lays in the size of the tensors or is it related to the way I am defining ops (given that the model parses, and optimizes successfully)?
  2. Would it be possible to allows multiple APUs for the same memory buffer?
  3. How would you resolve the error?